Research-based list of AI models, development tools, and automation resources. Use it to compare releases, pricing, benchmarks, and deployment options from official sources.
Looking for the best AI models? Check out our Top Models by Category table below for quick recommendations. For detailed analysis, explore the comprehensive sections on frontier models, open-source alternatives, and development tools.
- 138+ AI models from major providers (OpenAI, Anthropic, Google, Meta, xAI, Microsoft, and others)
- Latest pricing and performance benchmarks (latest data through 2026-07-11 19:14 UTC)
- GPT-5.6 reached general availability on 2026-07-09 across ChatGPT, Codex, and the OpenAI API; Sol, Terra, and Luna remain the three-tier family.
- Gemma 4 models were released through the Gemini API and AI Studio on 2026-07-06, including Gemma 4 31B and Gemma 4 26B-A4B.
- Claude Sonnet 5 launched 2026-06-30 00:00 UTC with 1M context, 128K output, $2.00 / $10.00 introductory pricing through 2026-08-31 00:00 UTC, and 92.4% SWE-bench Verified
- Claude Fable 5 and Mythos 5 access restored 2026-07-01 00:00 UTC after export controls were lifted; Fable 5 globally available while Mythos 5 remains restricted to Project Glasswing partners
- Cartesia Sonic-3.5 and Ink-2 are now official production speech models for low-latency voice agents
- MiniMax Speech 2.8 official page verified: native sound tags, high-fidelity cloning, and studio-grade multilingual TTS
- GPT-Live (OpenAI, 2026-07-08): full-duplex voice model family powering ChatGPT Voice; GPT-Live-1 default for paid tiers, GPT-Live-1 mini for Free; no API yet (
soon); GPT-5.5 backend delegates complex work. GPT-Realtime 2.1 / 2.1-mini shipped 2026-07-06 in the Realtime API with function calling, MCP, and SIP support - Kiro 1.0.52, Cursor Team Marketplace MCPs, Codex CLI 0.142.0, and Windsurf quota-based plans refreshed from official release notes
- Codex CLI 0.142.5 patched WebSocket trace logging, and GitHub Copilot added enterprise agent session streaming plus
GITHUB_TOKENsupport for Copilot CLI in GitHub Actions - Browser Use CLI 3.0, Browser Harness, and Vercel agent-browser refreshed in Browser Automation from official project docs and releases
- Vercel AI Gateway routing rules, eve Agent Runs in Vercel MCP/CLI, and Vercel Sandbox FUSE support added from official July 2026 changelogs
- Cloudflare Browser Run /json Quick Action, Jina v5 Omni embeddings, and official MCP Registry references refreshed from official docs
- MiniMax M2.7 self-evolving LLM with SWE-Pro 56.22%
- MAI-1, MAI-1 Plan, and MAI-Frontier Tunings launched July 2026 with Microsoft Frontier Tuning framework reaching GA status
- Cheapest verified direct APIs now tracked from official pricing pages: Groq Llama 3.1 8B Instant [$0.05/$0.08], Gemini 2.5 Flash-Lite [$0.10/$0.40], DeepSeek-V4-Flash [$0.14/$0.28], and OpenAI GPT-5.4 nano [$0.20/$1.25]
- Google AI Plus price drop: $7.99 → $4.99/month with storage doubled to 400GB (effective June 8, 2026)
- Self-hosting guides for open-source models
- Development tools for AI application building
| Category | #1 | #2 | #3 |
|---|---|---|---|
| Coding | Claude Opus 4.8 | GPT-5.5 | Claude Sonnet 5 |
| Reasoning | Gemini 3.1 Pro | GPT-5.5 Pro | Qwen3.7-Max |
| Open Source | GLM-5.2 | MiniMax M3 | DeepSeek-V4-Flash |
| Cost Efficiency | Llama 3.1 8B Instant | Gemini 2.5 Flash-Lite | DeepSeek-V4-Flash |
| Free & Budget | Gemini 2.5 Flash-Lite | Command A+ | Mistral Small 4 |
| Agentic Performance | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
| Context Window | Gemini 3 Flash (10M) | Llama 4 Scout (10M) | Claude Opus 4.8 (1M) |
- Models 🧠 - Comprehensive model comparisons and specifications
- Development Tools 🛠️ - IDEs, CLI tools, and API providers
- Automation 🤖 - Browser automation and agent platforms
- Guides 📚 - Getting started and selection guides
- AI Infrastructure 🏗️ - Models, frameworks, and deployment options
- Document Version: 3.46
- Last Updated: 2026-07-11 19:14 UTC
- Repository: github.com/ReadyPixels/AI_Models_Matrix
- License: CC BY-NC 4.0
- Ctrl+F - Use your browser's search to find specific models or topics
- Jump to sections using the table of contents above
- Model search - Use your browser's search to find specific models (e.g., "GPT", "Claude", "Gemini")
Comprehensive documentation of Large Language Models (LLMs), Small Language Models (SLMs), and specialized AI models available today.
State-of-the-art proprietary AI models with cutting-edge capabilities from leading AI labs.
| Model | Company | Context | GPQA Diamond | Arena Elo | SWE-bench Verified | AIME 2025 | Pricing | Verified |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 4.8 | Anthropic | 1M | 93.6% | 1506 | 88.6% | — | $5.00 / $25.00 | 2026-05-28 |
| Claude Sonnet 5 ⭐ | Anthropic | 1M | — | — | 92.4% | — | $2.00 / $10.00 (intro) | 2026-06-30 00:00 UTC ⭐ |
| Claude Opus 4.7 | Anthropic | 1M | 94.2% | 1505 | 87.6% | ~95% | $5.00 / $25.00 | 2026-04-26 |
| GPT-5.5 | OpenAI | 1M | 93.6% | 1506 | 88.5% | 99.9% | $5.00 / $30.00 | 2026-04-26 |
| GPT-5.5 Instant | OpenAI | 1M | 91.0% | — | 88.7% | — | $0.20 / $1.00 | 2026-05-05 |
| GPT-5.5 Pro | OpenAI | 1.05M | 94.2% | 1551 | 92.3% | 100% | $30.00 / $180.00 | 2026-05-08 |
| Gemini 3.1 Pro | 1M | 94.3% | 1505 | 80.6% | 100% | $2.00 / $12.00 | 2026-04-26 | |
| Claude Fable 5 |
Anthropic | 1M | 94.5% | 1510 | 95.0% | — | $10.00 / $50.00 | 2026-06-09 ⭐ |
| Claude Mythos 5 |
Anthropic | 1M | ~94.1% | — | 95.5% | — | $10.00 / $50.00 | 2026-06-09 ⭐ |
| NVIDIA Nemotron 3 Ultra | NVIDIA | 262K | 87.0% | — | — | — | Free (NVIDIA NIM) | 2026-06-05 |
| Kimi K2.7 Code ⭐ | Moonshot AI | 262K | 65.8% | — | 78.2% | 91.5% | $0.95 / $4.00 | 2026-06-12 ⭐ |
| GLM-5.2 ⭐ | Zhipu AI | 1M | ~91.2% | — | — | 99.2% (AIME 2026) | $1.40 / $4.40 | 2026-06-13 ⭐ |
| MAI-Thinking-1 ⭐ | Microsoft | 256K | 84.2% | — | — | 97% | TBD (Private Preview) | 2026-06-02 ⭐ |
| MAI-Code-1-Flash ⭐ | Microsoft | — | — | — | — | — | Free (Copilot) | 2026-06-02 ⭐ |
| MiniMax M3 ⭐ | MiniMax | 1M | — | — | 80.5% | — | $0.30 / $1.20 | 2026-06-01 ⭐ |
| MiniMax M2.7 ⭐ | MiniMax | 205K | — | — | 56.22% (SWE-Pro) | — | $0.30 / $1.20 | 2026-03-18 ⭐ |
| Qwen3.7-Plus ⭐ | Alibaba | 1M | — | — | — | — | $0.40 / $2.40 | 2026-06-03 ⭐ |
| Step 3.7 Flash ⭐ | StepFun | 1M | — | — | — | — | Free (limited) | 2026-05-28 ⭐ |
| GPT-5.6 Sol ⭐ | OpenAI | 1M+ | — | — | — | — | $5.00 / $30.00 | 2026-07-09 ⭐ |
| GPT-5.6 Terra ⭐ | OpenAI | 1M+ | — | — | — | — | $2.50 / $15.00 | 2026-07-09 ⭐ |
| GPT-5.6 Luna ⭐ | OpenAI | 1M+ | — | — | — | — | $1.00 / $6.00 | 2026-07-09 ⭐ |
⭐ Claude Fable 5 and Mythos 5 redeployed: Anthropic says export controls were lifted on 2026-06-30 00:00 UTC, and access to Fable 5 and Mythos 5 was restored on 2026-07-01 00:00 UTC. Fable 5 is available globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with cloud partner re-enablement in progress. Mythos 5 access is restored for approved US organizations and remains limited to Project Glasswing partners. Fable 5 has a 30-day data retention requirement, uses a different tokenizer (+30% token overhead vs Opus), and is unavailable for Zero Data Retention (ZDR) agreements. Source: 🔗
⭐ GPT-5.6 Sol/Terra/Luna (OpenAI, 2026-07-09): General availability release following the limited preview. Three-tier family: Sol ($5.00/$30.00), Terra ($2.50/$15.00), Luna ($1.00/$6.00) per 1M tokens. Source: 🔗
⚠️ GPT-5.5 Instant pricing correction: The API pricing for GPT-5.5 Instant is $0.20/$1.00 per 1M tokens (input/output), not $5.00/$30.00 which is the standard GPT-5.5 rate. GPT-5.5 Instant replaced GPT-5.3 Instant as the free-tier ChatGPT default on May 5, 2026.
Detailed technical specifications, pricing, and capabilities for frontier models and notable API models. Data as of 2026-07-03 21:56 UTC.
Maximum output tokens per single API request.
| Model | Max Output | Context Window | Notes |
|---|---|---|---|
| Claude Opus 4.6 | 128K (300K via beta) | 1M | Extended output via output-128k-2025-02-19 beta header |
| Claude Opus 4.7 | 128K (300K via beta) | 1M | Extended output via output-128k-2025-02-19 beta header |
| Claude Sonnet 4.6 | 64K | 1M | — |
| Claude Sonnet 5 ⭐ | 128K (300K via beta) | 1M | Released 2026-06-30 00:00 UTC; adaptive thinking on by default; batch API supports 300K with output-300k-2026-03-24 beta header |
| Claude Sonnet 4.5 | 64K | 200K | — |
| GPT-5.4 | 128K | 1.05M | — |
| GPT-5.4 mini | 128K | 400K | — |
| GPT-5.4 nano | 128K | 400K | — |
| GPT-5.3-Codex | 128K | 400K | — |
| GPT-5.5 Instant | 128K | 1M | Default ChatGPT model since 2026-05-05 |
| Gemini 3.1 Pro | 64K | 1M | — |
| Gemini 3.5 Flash | 64K | 1M | Released 2026-05-19 at Google I/O 2026; 4x faster than 3.1 Pro |
| Gemini 3 Pro | 64K | 2M | — |
| Gemini 3 Flash | 64K | 1M | — |
| Gemini 3.1 Flash-Lite | 32K | 1M | — |
| DeepSeek-V4-Flash | 384K | 1M | API model ID deepseek-v4-flash |
| DeepSeek-V4-Pro | 384K | 1M | API model ID deepseek-v4-pro; 75% discount made permanent 2026-05-31; standard price $0.435 / $0.87 |
| DeepSeek-V3.2 | 8K / 64K (reasoner) | 128K | Reasoner mode unlocks 64K output |
| Qwen3.5-Max | 65K | 1M | — |
| Qwen3.7-Max | 65K | 1M | Extended thinking by default; proprietary, API-only |
| GLM-5 | 128K | 200K | — |
| GLM-5.1 | 131K | 200K | — |
| MiniMax M3 | 512K | 1M (guaranteed 512K min) | Native multimodal (text/image/video); open weights pending ~2026-06-11 |
| MiniMax M2.7 | 262K | 205K | Self-evolution capabilities; Agent Teams native; SWE-Pro 56.22% |
| MiniMax-M2.5 | 131K | 1M | — |
| Kimi K2.6 | 262K | 262K | Max output = 262,144 tokens (MIT open-weight) |
| Kimi K2.7 Code | 262K | 262K | Coding-focused variant of K2.6; 30% fewer thinking tokens; Modified MIT |
| Step-3.5-Flash | 66K | 256K | — |
| Grok 4 | — | 256K | Not publicly specified |
| Grok 4 Fast | 30K | 128K | Now aliased to Grok 4.3; $0.20 / $0.50 |
| Grok 4.20 | 30K | 2M | All variants: $1.25 / $2.50 |
| Grok 4.3 | 30K | 1M | Always-on reasoning; released 2026-05-01 |
| Mistral Small 4 | 8K | 256K | 119B MoE; multimodal (text+image); $0.15 / $0.60 |
| Mistral Medium 3.5 | 8K | 256K | Released 2026-04-29; 128B dense; $1.50/$7.50 |
| Mistral Large 3 | 8K | 256K | — |
| Llama 4 Scout | 16K | 10M | — |
| Llama 4 Maverick | 16K | 1M | — |
| Claude Fable 5 | 128K | 1M | Mythos-class; $10.00 / $50.00; released 2026-06-09 |
| Claude Mythos 5 | 128K | 1M | Mythos-class (restricted); $10.00 / $50.00; released 2026-06-09 |
| GPT-4.1 ⭐ | 128K | 1M | Knowledge cutoff June 2024; $2.00 / $8.00; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04 |
| GPT-4.1 mini ⭐ | 128K | 1M | Knowledge cutoff June 2024; $0.40 / $1.60; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04 |
| GPT-4.1 nano ⭐ | 128K | 1M | Knowledge cutoff June 2024; $0.10 / $0.40; retired from ChatGPT Feb 13, 2026; API still available; released 2026-04 |
| GPT-5.6 Sol ⭐ | 128K | 1M+ | GA model; $5.00 / $30.00; released 2026-07-09 |
| GPT-5.6 Terra ⭐ | 128K | 1M+ | GA model; $2.50 / $15.00; released 2026-07-09 |
| GPT-5.6 Luna ⭐ | 128K | 1M+ | GA model; $1.00 / $6.00; released 2026-07-09 |
| Gemini 3.1 Flash Image ⭐ | — | 131K | Image + text + audio understanding; generation; $0.50/1M text+image input tokens; output images priced by resolution ($0.045/512px, $0.067/1Kpx, $0.101/2Kpx); GA June 2026 |
Discounted pricing tiers for high-volume usage. All prices in USD per million tokens.
| Model | Standard Input | Cached Input | Batch Discount | Notes |
|---|---|---|---|---|
| Claude Opus 4.7 | $5.00 | $0.50 (hit) / $6.25 (5m write) | 50% off | Batch: $2.50 in / $12.50 out |
| Claude Opus 4.8 | $5.00 | $0.50 (hit) / $6.25 (5m write) | 50% off | Same pricing as 4.7; fast mode $10/$50 |
| Claude Opus 4.6 | $5.00 | $0.50 (hit) / $6.25 (5m write) | 50% off | Batch: $2.50 in / $12.50 out |
| Claude Sonnet 4.6 | $3.00 | $0.30 (hit) / $3.75 (5m write) | 50% off | Batch: $1.50 in / $7.50 out |
| Claude Sonnet 5 ⭐ | $2.00 (intro) / $3.00 standard | $0.20 hit / $2.50 5m write (intro) | 50% off | Intro pricing through 2026-08-31 00:00 UTC: $2.00 / $10.00; standard from 2026-09-01 00:00 UTC: $3.00 / $15.00 |
| Claude Sonnet 4.5 | $3.00 | $0.30 (hit) / $3.75 (5m write) | 50% off | Batch: $1.50 in / $7.50 out |
| GPT-5.4 | $2.50 | $0.25 | 50% off | Data residency +10% |
| GPT-5.4 mini | $0.75 | $0.075 | 50% off | — |
| GPT-5.4 nano | $0.20 | $0.02 | 50% off | — |
| GPT-5.3-Codex | $1.75 | $0.175 | 50% off | — |
| GPT-5.5 Instant | $0.20 | $0.02 | 50% off | Free-tier ChatGPT default since 2026-05-05; API pricing $0.20/$1.00 |
| Gemini 3.1 Pro | $2.00 | $0.20–$0.40 + $4.50/hr storage | 50% off | Tiered by input length |
| Gemini 3.5 Flash | $1.50 | $0.15 | 50% off | Released 2026-05-19; 4x faster than 3.1 Pro |
| Gemini 3 Flash | $0.50 | $0.05 + $1.00/hr storage | 50% off | — |
| Gemini 3.1 Flash-Lite | $0.25 | $0.025 + $0.25/hr storage | 50% off | Cached input $0.025/1M (Vertex standard tier, ≤200K ctx) |
| DeepSeek-V4-Flash | $0.14 | $0.0028 (hit) | — | Cache-hit pricing from 2026-04-26 12:15 UTC |
| DeepSeek-V4-Pro | $0.435 | $0.003625 (hit) | — | 75% discount made permanent (2026-05-31); list price now $0.435 / $0.87 |
| DeepSeek-V3.2 | $0.28 | $0.028 | — | No formal batch API |
| Qwen3.5-Max | $0.40 | Available | 50% off | — |
| Qwen3.7-Max | $2.50 | $0.25 | — | 90% cached discount; proprietary API-only |
| GLM-5 / GLM-5.1 | $1.00 | $0.20 | — | — |
| Grok 4 | $3.00 | $0.75 | — | — |
| Grok 4 Fast | $0.20 | $0.05 | — | Aliased to Grok 4.3; output $0.50/1M |
| Grok 4.20 | $1.25 | $0.20 | — | All variants: reasoning, non-reasoning, multi-agent |
| Grok 4.3 | $1.25 | $0.20 | — | ~38% cheaper input than 4.20; released 2026-05-01 |
| Mistral Small 4 | $0.15 | $0.015 (estimated) | — | 119B MoE; multimodal; released 2026-03-16 |
| Command A+ ⭐ | $0.40 | — | — | 25B active (218B total MoE); faster than GPT-5.4 nano, Claude Haiku, Grok 4.3; released 2026-05 |
| Mistral Large 3 | $0.50 | $0.05 (cached) | — | Cached input $0.05/1M; 256K context |
| Step-3.5-Flash | $0.10 | — | — | — |
| Claude Fable 5 | $10.00 | $1.00 (hit) / $12.50 (5m write) | 50% off | Batch: $5.00 in / $25.00 out; fast mode $20/$100 |
| Claude Mythos 5 | $10.00 | $1.00 (hit) / $12.50 (5m write) | 50% off | Same pricing as Fable 5; restricted access via Project Glasswing |
| MiniMax M2.7 | $0.30 | — | — | Self-evolution capabilities; Agent Teams native; SWE-Pro 56.22%; 205K context; released 2026-03-18 |
| MiniMax M3 (≤512K) | $0.30 | $0.06 (hit) | — | Extended context (>512K): $0.60/$2.40; released 2026-06-01 |
| Qwen3.7-Plus | $0.40 | — | — | Released 2026-06-03; 1M context |
| Step 3.7 Flash | Free | — | — | Free tier available; 1M context; released 2026-05-28 |
| NVIDIA Nemotron 3 Ultra | Free | — | — | Free via NVIDIA NIM; 262K context |
| GPT-4.1 ⭐ | $2.00 | $0.20 | — | 1M context; knowledge cutoff June 2024; retired from ChatGPT Feb 13, 2026; API still available |
| GPT-4.1 mini ⭐ | $0.40 | $0.04 | — | 1M context; retired from ChatGPT Feb 13, 2026; API still available |
| GPT-4.1 nano ⭐ | $0.10 | $0.01 | — | 1M context; retired from ChatGPT Feb 13, 2026; API still available; OpenAI's cheapest model |
| GPT-5.6 Sol ⭐ | $5.00 | $0.50 (90% cached) / $6.25 (cache write) | 50% off | GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09 |
| GPT-5.6 Terra ⭐ | $2.50 | $0.25 (90% cached) / $3.125 (cache write) | 50% off | GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09 |
| GPT-5.6 Luna ⭐ | $1.00 | $0.10 (90% cached) / $1.25 (cache write) | 50% off | GA; cache write 1.25x input; cached read 90% discount; released 2026-07-09 |
Output throughput and time-to-first-token from Artificial Analysis and provider benchmarks.
| Model | Output Speed (tok/s) | TTFT | Notes |
|---|---|---|---|
| Gemini 3.1 Flash-Lite | ~250 | ~2.1s | Fastest budget Google model |
| Step-3.5-Flash | 85–350 | — | Variable by provider; peak ~350 tok/s |
| Gemini 3 Flash | ~193 | ~4.16s | — |
| MiniMax-M2.5 Lightning | ~100 | — | Faster tier |
| GPT-5.3-Codex | ~86 | ~77.86s | High TTFT due to extended reasoning |
| Grok 4 | ~56 | ~8.96s | — |
| MiniMax-M2.5 Standard | ~50 | — | — |
Most frontier models (Claude Opus/Sonnet 4.6, GPT-5.4, Gemini 3.1 Pro, etc.) have not yet been benchmarked on Artificial Analysis as of April 2026. Anthropic notes that fast mode on Claude Opus 4.7 is deprecated and scheduled for removal on 2026-07-24 00:00 UTC; use Claude Opus 4.8 for current Opus fast-mode workflows. Source: 🔗
Knowledge cutoff dates — the point after which a model has no training data.
| Model | Training Cutoff | Notes |
|---|---|---|
| Claude Sonnet 4.6 | Jan 2026 | Most recent cutoff among frontier models |
| Claude Sonnet 5 ⭐ | Jan 2026 | Reliable knowledge cutoff Jan 2026; released 2026-06-30 00:00 UTC |
| Claude Opus 4.6 | Aug 2025 | Reliable knowledge: May 2025 |
| GPT-5.4 / mini / nano | Aug 31, 2025 | — |
| GPT-5.3-Codex | Aug 31, 2025 | — |
| Grok 4 Fast | Jul 2025 | — |
| Grok 4.20 / 4.3 | Jul 2025 | Approximate; not publicly disclosed |
| DeepSeek-V4 (Flash/Pro) | May 2025 | — |
| Gemini 3.1 Flash-Lite | Jan 2025 | — |
| Gemini 3.1 Pro / 3 Pro / 3 Flash | Jan 2025 | — |
| Grok 4 | ~Nov–Dec 2024 | Approximate |
| DeepSeek-V3.2 | Jul 2024 | — |
| Llama 4 Scout / Maverick | Aug 2024 | — |
| DeepSeek-R1 | ~Oct 2023 | Based on base model |
| GPT-4.1 / mini / nano | June 2024 | Retired from ChatGPT Feb 13, 2026; API still available |
| GPT-4.5 | June 2024 | Retired from API July 2025; retiring from ChatGPT 2026-06-27 |
Models not listed (Qwen, GLM, MiniMax, Kimi, Step, Mistral): training cutoff not publicly disclosed.
| Model | Languages | Details |
|---|---|---|
| Qwen3.5-Max | 201 | Largest language coverage |
| Llama 4 Scout | 200 | Pre-training languages |
| Qwen3-Max-Thinking | 119 | Qwen3 series |
| Gemini 3 Flash | 100 | 91.8% MMMLU score across 100 languages |
| Gemini 3.1 Pro / 3 Pro | 100+ | — |
| Gemini 3.1 Flash-Lite | 100 | 91.3% MMMLU score |
| Llama 4 Maverick | 12 | Output languages |
| Claude (all) | Many | English-optimized; broad multilingual; Sonnet 5 supports multilingual capabilities and vision |
| GPT-5.4 (all) | Many | Broad multilingual coverage |
| DeepSeek (all) | Many | Chinese + English focused |
| Grok (all) | Many | — |
| GLM-5 / GLM-5.1 | Many | 28.5T token training data |
All frontier models support structured JSON output and function/tool calling except where noted.
| Capability | Supported Models | Not Supported |
|---|---|---|
| Structured Output (JSON mode) | All models listed in Frontier table | Gemini 3 Deep Think (no API) |
| Function Calling / Tool Use | All models listed in Frontier table | Gemini 3 Deep Think (no API) |
Gemini 3 Deep Think is available only via Gemini's in-app Think mode — no API access for structured output or function calling.
| Provider | API Availability | Cloud Partners | Notes |
|---|---|---|---|
| Anthropic | Global | AWS Bedrock, GCP Vertex AI | US-only inference at 1.1x via inference_geo |
| OpenAI | Global | Azure OpenAI | Data residency endpoints +10% (post-3/5/26) |
| Global | Google AI Studio, Vertex AI | Some regional restrictions per Google terms | |
| DeepSeek | Global | Azure (R1 only, select regions) | China-based servers |
| Alibaba (Qwen) | Global | Alibaba Cloud Model Studio | China-based; globally accessible |
| Zhipu AI (GLM) | Global | Z.AI API | MIT license enables self-hosting anywhere |
| MiniMax | Global | MiniMax API | — |
| Moonshot AI (Kimi) | Global | platform.kimi.ai | MIT open-weight |
| xAI (Grok 4 / 4.20 / 4.3 / 4 Fast) | US-focused | Oracle OCI (East/Midwest/West), Amazon Bedrock (June 2026) | Major price cuts: Grok 4.20 now $1.25/$2.50; Grok 4 Fast aliased to 4.3; Grok 4.3 on Bedrock at $1.25/$2.50 |
| Mistral | Global | Azure AI Foundry, AWS, GCP | — |
| Meta (Llama) | Global (self-host) | All major cloud providers | Llama 4 Community License |
| StepFun | Global | HuggingFace | Apache 2.0 open-source |
Self-hostable models with permissive licenses or open weights for privacy, cost control, and customization.
| Model | Company | Params | Context | License |
|---|---|---|---|---|
| DeepSeek-V4-Flash | DeepSeek | 1.6T / 49B active (MoE) | 1M | Open Weight |
| Qwen3.5-Max | Alibaba | 397B / 17B active (MoE) | 262K | Apache 2.0 |
| Qwen3-Max-Thinking | Alibaba | 1T+ | 128K | Apache 2.0 |
| Qwen3.6-27B | Alibaba | 27B dense | 262K | Apache 2.0 |
| Qwen3.5-122B | Alibaba | 397B / 17B active (MoE) | 262K | Apache 2.0 |
| Qwen3.5-27B | Alibaba | 27B dense | 262K | Apache 2.0 |
| Mistral Small 4 | Mistral AI | 119B / 6B active (MoE, 128 experts) | 256K | Apache 2.0 |
| Mistral Large 3 | Mistral AI | 675B / 41B active (MoE) | 256K | Apache 2.0 |
| Command A+ | Cohere | 25B active (218B total MoE) | 128K | Apache 2.0 |
| Llama 4 Scout | Meta | 109B | 10M | Community |
| Llama 4 Maverick | Meta | 400B | 128K | Community |
| GPT-OSS-120B | OpenAI | 117B | 128K | Apache 2.0 |
| GPT-OSS-20B | OpenAI | 21B | 128K | Apache 2.0 |
| Qwen3-Coder | Alibaba | 480B | 262K | Apache 2.0 |
| GLM-5.1 | Zhipu AI | 754B / 40B active (MoE) | 200K | MIT |
| GLM-4.7 | Zhipu AI | 400B+ MoE | 205K | Open Weight |
| Gemma 4 31B | 31B dense | 256K | Apache 2.0 | |
| Gemma 4 26B-A4B | 26B MoE (4B active) | 256K | Apache 2.0 | |
| Gemma 4 E4B | 4B dense | 256K | Apache 2.0 | |
| Gemma 4 E2B | 2B dense | 256K | Apache 2.0 | |
| Gemma 4 12B | 12B dense | 256K | Apache 2.0 | |
| Qwen3-Coder 7B | Alibaba | 7B dense | 128K | Apache 2.0 |
| Qwen 2.5 Coder 32B | Alibaba | 32B dense | 128K | Apache 2.0 |
| DeepSeek Coder-V2 | DeepSeek | 236B / 2.4B active | 128K | MIT |
| Step-3.5-flash | StepFun | 196B / 11B active (MoE) | 256K | Open Weight |
| Yi-Coder | 01.AI | 9B/1.5B | 128K | Apache 2.0 |
| Lizzy-7B | Flower Labs | 7B | — | MIT |
| MiMo-V2.5 | Xiaomi | 309B / 15B active | 1M | MIT |
| MiMo-V2.5-Pro | Xiaomi | 1.02T / 42B active | 1M | MIT |
| NVIDIA Nemotron 3 Ultra | NVIDIA | 550B / 55B active (MoE) | 262K (BF16) / 1M (NVFP4 Blackwell) | Open Weight (NVIDIA) |
| MiniMax M3 ⭐ | MiniMax | ~428B / ~23B active (MoE) | 1M | MiniMax Community |
| MiniMax M2 ⭐ | MiniMax | 230B / 10B active (MoE) | 1M | Modified MIT |
| MiniMax M2.1 ⭐ | MiniMax | 229B / 10B active (MoE) | 1M | Apache 2.0 |
| GLM-5.2 ⭐ | Zhipu AI | 753B / 40B active (MoE) | 1M | MIT |
| Kimi K2.7 Code ⭐ | Moonshot AI | 1T / 32B active (MoE) | 262K | Modified MIT |
| MiniMax M2.7 ⭐ | MiniMax | ~400B+ (MoE) | 205K | MiniMax Community |
| Kimi K2.6 | Moonshot AI | 1T / 32B active (MoE) | 262K | Modified MIT |
| Ornith-1.0 ⭐ | DeepReinforce | 9B / 31B / 35B-MoE / 397B-MoE | — | MIT |
Local Inference Tools:
- Ollama - Easy local deployment
- LM Studio - User-friendly GUI
- llama.cpp - Efficient CPU inference
- vLLM - High-throughput serving
- SGLang - Structured generation
Cloud Deployment:
- Hugging Face Inference - Managed deployment
- AWS SageMaker - Full control
- Google Cloud Vertex - Integrated
- RunPod - GPU rental
Specialized AI models optimized for software development tasks.
| Rank | Model | Company | SWE-bench Verified |
|---|---|---|---|
| 🥇 #1 | Claude Mythos 5 |
Anthropic | 95.5% |
| 🥈 #2 | Claude Fable 5 |
Anthropic | 95.0% |
| 🥉 #3 | Claude Mythos Preview | Anthropic | 93.9% |
| #4 | Claude Sonnet 5 ⭐ | Anthropic | 92.4% |
| #5 | GPT-5.5 Pro | OpenAI | 92.3% |
| #6 | GPT-5.5 | OpenAI | 88.7% |
| #7 | Claude Opus 4.8 | Anthropic | 88.6% |
| #8 | Claude Opus 4.7 | Anthropic | 87.6% |
| #9 | GPT-5.3-Codex | OpenAI | 85.0% |
| #10 | Ornith-1.0-397B ⭐ | DeepReinforce | 82.4% |
| #11 | Claude Opus 4.5 | Anthropic | 80.9% |
| #12 | Claude Opus 4.6 | Anthropic | 80.8% |
| #13 | DeepSeek-V4-Pro (Max) | DeepSeek | 80.6% |
| #13 | Gemini 3.1 Pro | 80.6% | |
| #14 | MiniMax M3 ⭐ | MiniMax | 80.5% |
| #14 | Qwen3.7-Max | Alibaba | 80.4% |
| #15 | Kimi K2.6 | Moonshot AI | 80.2% |
| #16 | MiniMax-M2.5 | MiniMax | 80.2% |
| #17 | GPT-5.2 | OpenAI | 80.0% |
| #18 | Claude Sonnet 4.6 | Anthropic | 79.6% |
| #19 | DeepSeek-V4-Flash (Max) | DeepSeek | 79.0% |
| #20 | Qwen3.6 Plus | Alibaba | 78.8% |
| #21 | Kimi K2.7 Code ⭐ | Moonshot AI | 78.2% |
| #22 | Gemini 3 Flash | 78.0% | |
| #22 | MiMo-V2-Pro | Xiaomi | 78.0% |
| #24 | GLM-5 | Zhipu AI | 77.8% |
| #25 | Mistral Medium 3.5 | Mistral AI | 77.6% |
| #26 | Claude Sonnet 4.5 | Anthropic | 77.2% |
| Model | Developer | Pricing | Best For |
|---|---|---|---|
| Claude Fable 5 |
Anthropic | $10.00 / $50.00 per 1M | Mythos-class coding; SWE-bench 95.0%; restored globally 2026-07-01 00:00 UTC with stronger safeguards |
| Claude Opus 4.8 | Anthropic | $5.00 / $25.00 per 1M | Agentic coding, complex tasks, best available Opus |
| Claude Sonnet 5 ⭐ | Anthropic | $2.00 / $10.00 per 1M intro through 2026-08-31 00:00 UTC | Agentic coding at Sonnet latency and price; 92.4% SWE-bench Verified; 63.2% SWE-bench Pro |
| Kimi K2.7 Code ⭐ | Moonshot AI | $0.95 / $4.00 per 1M | Open-weight coding agent; 30% fewer thinking tokens vs K2.6; now available in GitHub Copilot model picker (July 1, 2026) |
| Claude Opus 4.6 | Anthropic | $5.00 / $25.00 per 1M | Agentic coding, complex tasks |
| GPT-5.5 Pro | OpenAI | $30.00 / $180.00 per 1M | Highest benchmark coding |
| GPT-5.3-Codex | OpenAI | $1.75 / $14.00 per 1M | Agentic coding, 7+ hour autonomy |
| Claude Haiku 4.5 | Anthropic | $1.00 / $5.00 per 1M | Low-latency coding, sub-agents, computer use |
| GLM-5-Code | Zhipu AI | $1.20 / $5.00 per 1M | Code generation, refactoring |
| MiniMax M3 | MiniMax | $0.30 / $1.20 per 1M | Frontier coding + 1M context + native multimodal; open weights pending |
| MiniMax M2 ⭐ | MiniMax | $0.30 / $1.20 per 1M | Open-weight MoE (230B/10B active); 1M context; 10x cheaper than Sonnet; released 2026-07-02 |
| MiniMax M2.1 ⭐ | MiniMax | $0.30 / $1.20 per 1M | Open-weight MoE (229B/10B active); 1M context; Apache 2.0; multilingual coding; released 2026-07-02 |
| MiniMax M2.7 ⭐ | MiniMax | $0.30 / $1.20 per 1M | Self-evolution; Agent Teams native; SWE-Pro 56.22%; 205K context; released 2026-03-18 |
| MiniMax-M2.5 | MiniMax | $0.30 / $1.20 per 1M | Code generation, refactoring |
| Claude Sonnet 4.5 | Anthropic | $3.00 / $15.00 per 1M | Code review, refactoring |
| Mistral Small 4 | Mistral AI | $0.15 / $0.60 per 1M | Unified reasoning + coding + multimodal; open-weight |
| Codestral | Mistral AI | $0.30 / $0.90 | Real-time completion |
| North Mini Code | Cohere | Free (OpenRouter) / API | 30B MoE (3B active); Apache 2.0; released 2026-06-09 |
| Qwen3.7-Max | Alibaba | $2.50 / $7.50 per 1M | Agentic coding, long-horizon autonomy, reasoning |
| Grok 4.3 | xAI | $1.25 / $2.50 per 1M | Flagship reasoning, lowest hallucination |
| MAI-Code-1-Flash | Microsoft | $0.75 / $4.50 per 1M | GitHub Copilot integration, fast everyday coding |
| Model | Developer | License | Hardware |
|---|---|---|---|
| Kimi K2.7 Code ⭐ | Moonshot AI | Modified MIT | 80-160 GB VRAM |
| GLM-5.2 ⭐ | Zhipu AI | MIT | 80-160 GB VRAM |
| MiniMax M2 ⭐ | MiniMax | Apache 2.0 | 48-80 GB VRAM |
| MiniMax M2.1 ⭐ | MiniMax | Apache 2.0 | 48-80 GB VRAM |
| MiniMax M3 ⭐ | MiniMax | MIT | 80-160 GB VRAM |
| MiniMax M2.7 ⭐ | MiniMax | MiniMax Community | 80-160 GB VRAM |
| Ornith-1.0 ⭐ | DeepReinforce | MIT | 160-320 GB VRAM |
| GPT-OSS-120B | OpenAI | Apache 2.0 | 80-160 GB VRAM |
| Qwen3-Coder | Alibaba | Apache 2.0 | 160-320 GB VRAM |
| DeepSeek-Coder-V2 | DeepSeek | MIT | 48-80 GB VRAM |
| GLM-5.1 | Zhipu AI | MIT | 80-160 GB VRAM |
| Phi-4 | Microsoft | MIT | 24-48 GB VRAM |
| Qwen3-Coder 7B | Alibaba | Apache 2.0 | 24-48 GB VRAM |
Models optimized for step-by-step reasoning, mathematical problem-solving, and complex logical inference.
| Rank | Model | AIME 2025 | ARC-AGI-2 | Notes |
|---|---|---|---|---|
| 🥇 #1 | MAI-Thinking-1 ⭐ | 97% | — | Microsoft's first in-house reasoning model; private preview |
| #2 | Kimi K2.5 (Reasoning) | 96.1% | — | Moonshot AI reasoning variant |
| #2 | Kimi K2.5 | 96.1% | — | Open-weight reasoning |
| #4 | GLM-4.7 | 95.7% | — | Zhipu AI open-weight |
| #5 | MiMo-V2-Flash | 94.1% | — | Xiaomi open-weight |
| #6 | Claude Sonnet 4.5 | 87% | — | — |
| #7 | Exaone 4.0 32B | 85.3% | — | LG AI Research |
| #8 | Nemotron 3 Nano Omni 30B | 82.1% | — | NVIDIA |
| #9 | LFM2.5-8B-A1B | 42.5% | — | LiquidAI |
| #10 | MiniCPM5-1B | 40.4% | — | OpenBMB |
| #12 | Step-3.5-Flash | 97.3% | — | Best efficiency ratio |
| #12 | GLM-5.2 ⭐ | 99.2% (AIME 2026) | — | Zhipu AI open-weight; leads AIME 2026 leaderboard |
| #13 | Kimi K2.6 | 96.4% | — | Strong multimodal reasoning |
| #14 | GLM-4.7 | 95.7% | — | Zhipu AI strong performer |
| #15 | Claude Sonnet 4.6 | ~95%–95.6% | 58.3% | Near-Opus performance |
| #16 | GLM-5 | 92.7% | — | Thinking mode |
| #17 | o4-mini | 92.7% | — | Efficient OpenAI reasoning |
| Model | Type | Context | Pricing |
|---|---|---|---|
| Gemini 3 Deep Think | Reasoning | 1M+ | Ultra subscription |
| Qwen3-Max-Thinking | Reasoning/Coding | 128K | $1.20 / $6.00 |
| o3 / o1-Pro | Reasoning | 128K | $2-150 / $8-600 |
| GPT-5.5 Pro | Reasoning | 1.05M | $30.00 / $180.00 |
| Gemini 3 Pro | General/Multimodal | 1M+ | $2.00 / $12.00 |
| DeepSeek-R1 | Reasoning | 128K | $0.50 / $2.15 |
| Claude Sonnet 4.5 | Hybrid | 200K | $3.00 / $15.00 |
| Claude Sonnet 5 ⭐ | Hybrid | 1M | $2.00 / $10.00 intro; $3.00 / $15.00 standard |
| GPT-Rosalind | Life Sciences Reasoning | 128K | Pay-per-token (Research Preview) |
| MAI-Thinking-1 | Reasoning | 256K | TBD (Private Preview) |
- Mathematical Problem Solving: Qwen3-Max-Thinking, GPT-5.5 Pro, Gemini 3 Pro
- Scientific Analysis: Claude Sonnet 5, Claude Opus 4.6, GPT-5.5, Gemini 3 Pro
- Strategic Planning: o3/o1-Pro, Claude Sonnet 4.5, DeepSeek-R1
- Code Debugging: Claude Sonnet 5, Claude Sonnet 4.5, GPT-5.3-Codex, DeepSeek-V3.2
Models capable of processing and generating multiple types of content: text, images, audio, and video.
| Model | Developer | Context | Key Features |
|---|---|---|---|
| GPT-5.4 | OpenAI | 1M | Unified multimodal, audio |
| Gemini 3 Pro | 1M+ | Native multimodal, video | |
| Claude Sonnet 4.5 | Anthropic | 200K | Document understanding |
| Claude Sonnet 5 ⭐ | Anthropic | 1M | Text and image input, 128K output, agentic coding and tool use |
| Llama 4 Maverick | Meta | 128K | Open multimodal |
| Gemini Omni Flash | — | Native multimodal (text+image+audio+video input, video generation), any-to-any creation | |
| Nemotron 3 Nano Omni | NVIDIA | 30B (3B active) | Vision, audio, language unified, 9x throughput |
| MiniMax M3 | MiniMax | 1M | First open-weight model: frontier coding + 1M context + native image & video understanding |
| Model | MMMU / MMMU-Pro | MathVista | DocVQA |
|---|---|---|---|
| Gemini 3.1 Pro | 95% (MMMU-Pro) | — | — |
| GPT-5.4 | 94% (MMMU-Pro) | — | — |
| Gemini 3 Pro | 81% (MMMU-Pro) | — | — |
| Gemini 3 Flash | 80% (MMMU-Pro) | — | — |
| Claude Sonnet 4.5 | 77.8% (MMMU) | — | — |
| Llama 4 Maverick | 73.4% (MMMU) | — | — |
| Model | Speech-to-Text | Text-to-Speech | Video Input |
|---|---|---|---|
| Gemini 3 Pro | ✅ | ✅ | ✅ |
| GPT-5 | ✅ | ✅ | |
| Whisper v3 | ✅ | ❌ | ✅ |
| Model | Developer | License | Best For |
|---|---|---|---|
| Gemini 3.1 Flash-Lite Image ⭐ | Proprietary | Gemini image model card updated 2026-06-30 00:00 UTC; image + text + audio understanding and generation; output images priced by resolution | |
| Nano Banana 2 Lite ⭐ | Proprietary | Fastest Gemini Image model; 4s generation; GA 2026-06-30; $0.034 per 1K images; rolling out to Gemini app, Search, NotebookLM, Photos, Stitch, Flow, Ads | |
| GPT Image 2 | OpenAI | Proprietary | ChatGPT Images 2.0; improved text rendering, multilingual support, advanced visual reasoning; released 2026-04-21; API pricing varies by quality/resolution |
| MAI-Image-2-Efficient | Microsoft | Proprietary | Production-ready quality, 41% lower cost |
| Flux 2 | Black Forest Labs | Apache 2.0 (Dev); Proprietary (Pro) | Released Nov 2025; exceptional photorealism + natural language; Flux 2 Pro/Dev/Schnell variants |
| Flux.1 | Black Forest Labs | Apache 2.0 | High-fidelity art (original Flux family) |
| Stable Diffusion 3.5 | Stability AI | Community License | Fine-tuning; MMDiT architecture, 2.5B params |
| GLM-Image | Zhipu AI (Z.ai) | API | Fast image generation |
| CogView-4 | Zhipu AI (Z.ai) | API | Creative image generation |
| Firefly AI Assistant | Adobe | Public Beta (2026-04-27) | Creative agent, 60+ tools, Photoshop/Premiere integration |
Comprehensive hardware specifications for self-hosting AI models.
| Model | Params | Q4 Size | Min VRAM | Rec VRAM | Min RAM |
|---|---|---|---|---|---|
| Phi-4 | 14B | 8 GB | 24 GB | 48 GB | 32 GB |
| GPT-OSS-20B | 21B | 12 GB | 24 GB | 48 GB | 32 GB |
| Llama 4 Scout | 109B | 66 GB | 48 GB | 80 GB | 96 GB |
| GPT-OSS-120B | 117B | 70 GB | 80 GB | 160 GB | 128 GB |
| DeepSeek-Coder-V2 | 236B | 143 GB | 48 GB | 80 GB | 192 GB |
| Llama 4 Maverick | 400B | 242 GB | 160 GB | 320 GB | 320 GB |
| DeepSeek-V4 family | 671B | 404 GB | 80 GB | 320 GB | 512 GB |
| Qwen3-Max-Thinking | 1T+ | 600+ GB | 160 GB | 640 GB | 768 GB |
Consumer/Entry Level (24-48 GB VRAM):
- Phi-4, GPT-OSS-20B, Yi-Coder, Qwen2.5-Coder
- Recommended GPUs: RTX 3090 (24GB), RTX 4090 (24GB)
Professional (80-160 GB VRAM):
- Llama 4 Scout, GPT-OSS-120B, DeepSeek-Coder-V2
- Recommended GPUs: A100 80GB, 2x A100 40GB
Enterprise (320+ GB VRAM):
- Llama 4 Maverick, GLM-4.7, DeepSeek-V4 family, Qwen3-Max-Thinking
- Recommended GPUs: 4x A100 80GB, 8x A100 80GB
| Level | Bits | Size vs FP16 | Quality | Use Case |
|---|---|---|---|---|
| FP16/BF16 | 16 | 100% | Best | Training |
| Q8_0 | 8 | ~50% | Excellent | High-quality inference |
| Q4_K_M | 4 | ~25% | Good | Recommended for deployment |
| Q3_K_M | 3 | ~19% | Fair | Limited resources |
Detailed benchmark scores across all major evaluations. Scores are percentages (%) unless noted. Arena Elo scores are integers. — = not publicly reported. Data as of 2026-07-03 21:56 UTC.
| Model | GPQA Diamond | MMLU-Pro | Arena Elo (Text) | HLE | SWE-bench Verified | SWE-bench Pro | LiveCodeBench | AIME 2025 | ARC-AGI-2 | MMMU-Pro | IFEval | FrontierMath |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 4.6 | 91.3% | — | 1500 | 40.0–53.0% | 80.8% | — | — | 99.8% | 68.8% | — | — | — |
| Claude Opus 4.8 | 93.6% | — | — | — | 88.6% | 69.2% | — | — | — | — | — | — |
| Claude Mythos Preview | 94.5% | — | — | — | 93.9% | — | — | — | — | — | — | — |
| GPT-5.5 | 93.6% | — | 1495 | 42.1–55.0% | 88.5% | — | — | 99.9% | 71.2% | — | — | 52% |
| GPT-5.5 Instant | 91.0% | — | — | — | 88.7% | — | — | — | — | — | — | — |
| NVIDIA Nemotron 3 Ultra | — | — | — | — | — | — | — | — | — | — | — | — |
| Ornith-1.0-397B ⭐ | — | — | — | — | 82.4% | — | — | — | — | — | — | — |
| GPT-5.5 Pro | 95.1% | 96% | 1520 | 48.5–62.0% | 92.3% | — | — | 100% | 78.5% | — | 97% | 58% |
| Claude Sonnet 4.6 | 89.9% | — | ~1438 | 33.2–49.0% | 79.6% | — | — | ~95% | 58.3% | — | — | — |
| Claude Sonnet 5 ⭐ | — | — | — | 43.2% no tools / 57.4% tools | 92.4% | 63.2% | — | — | — | — | — | — |
| Claude Sonnet 4.5 | 83.4% | 88.0% | — | — | 77.2% | — | — | 87–100% | — | — | — | — |
| Claude Fable 5 | 94.5% | — | — | — | 95.0% | 80.3% | 29.3% (FrontierCode) | — | — | — | — | — |
| Claude Mythos 5 | ~94.1% | — | — | — | 95.5% | — | — | — | — | — | — | — |
| Mistral Small 4 | ~71.2% | — | — | — | — | — | — | — | — | — | — | — |
| Mistral Medium 3.5 | — | — | — | — | 77.6% | — | — | 86.3% | — | — | — | — |
| GPT-5.4 | 92.0% | 94% | 1484 | 36.6–41.6% | ~80% | 57.7% | 84–88% | 88% | 73.3% | 94% | — | 50% (Pro) |
| GPT-5.4 mini | 87.5% | — | — | — | — | 54.4% | — | — | — | — | — | — |
| GPT-5.3-Codex | 91.5% | — | — | — | — | 56.8% | 85% | — | — | — | — | — |
| GPT-5.2 | 92.4% | — | 1479 | 35.2% | 80.0% | 55.6% | — | 100% | 52.9% | — | 95.6% | ~40.3% |
| Gemini 3.1 Pro | 94.3% | 92% | 1494 | 44.4–51.4% | 80.6% | 54.2–72% | 71% | 100% | 77.1% | 95% | 95% | — |
| DeepSeek-V4-Pro (Max) | 90.1% | — | — | — | 80.6% | 55.4% | — | — | — | — | — | — |
| Gemini 3.5 Flash | ~90.4% | — | — | — | 55.1% | — | — | 72.1% | 83.6% | 84.2% | 40.2% | — |
| Gemini 3 Pro | 91.9–93.8% | 83% | 1486 | 37.5% | 76.2% | 43.3% | 49% | 98–100% | 31.1–45.1% | 81% | 88% | 38% |
| Gemini 3 Flash | 90.4% | 72% | 1474 | 33.7% | 78.0% | 44% | — | — | — | 80% | 85% | — |
| Gemini 3 Deep Think | ~97% | 81% | — | 48.4% | ~58% | 63% | 58% | — | 84.6% | — | — | — |
| DeepSeek-V3.2 | 87.1% | 85.0% | — | 25.1% | 67.8% | — | — | 89.3% | — | — | — | — |
| DeepSeek-V4-Flash (Max) | 88.1% | — | — | — | 79.0% | 52.6% | — | — | — | — | — | — |
| DeepSeek-R1 | 71.5% | 84.0% | — | 8.5% | 49.2% | — | 63.5% | 70.0% | — | — | — | — |
| Qwen3.5-Max | 89.3% | — | — | — | 76.4% | — | — | 91.3% | — | 79% | — | — |
| Qwen3.7-Max | 92.4% | — | — | 41.4% | 80.4% | — | — | — | — | — | — | — |
| Qwen3.6 Plus | 90.4% | — | — | — | 78.8% | — | — | — | — | — | — | — |
| Qwen3-Max-Thinking | 86.1% | — | — | 26.2% | — | — | — | — | — | — | — | — |
| GLM-5 | 82.0% | — | ~1451 | 10.4% | 77.8% | — | — | 92.7% | — | — | — | — |
| MiMo-V2-Pro | 87.0% | — | — | — | 78.0% | — | — | — | — | — | — | — |
| GLM-5.1 | — | — | — | — | ~80.4% (est.) | — | — | — | — | — | — | — |
| Kimi K2.6 | 90.5% | 87.1% | — | 31.5–50.2% | 80.2% | — | 85.0% | 96.4% | — | 78.5% | — | — |
| GLM-5.2 ⭐ | ~91.2% | — | — | — | — | 62.1% | 76.8% (MCP-Atlas) | 99.2% | — | — | — | — |
| MiniMax M3 | ~92.9% | — | — | — | — | — | — | — | — | — | — | — |
| MiniMax-M2.5 | 85.2% | — | — | — | 80.2% | 55.4% | — | 86.3% | — | — | — | — |
| Step-3.5-Flash | 83.1% | — | — | — | 74.4% | — | 86.4% | 97.3% | — | — | — | — |
| Grok 4 | ~91.5% | 91.5% | ~1493 | 50.7% | — | — | — | 100% | — | — | — | — |
| Grok 4.20 | — | — | — | — | — | — | — | — | — | — | — | — |
| Grok 4.3 | — | — | — | — | — | — | — | — | — | — | — | — |
| Llama 4 Maverick | 69.8% | 80.5% | — | — | — | — | 43.4% | — | — | — | — | — |
| Llama 4 Scout | 57.2% | 74.3% | — | — | — | — | 32.8% | — | — | — | — | — |
FrontierMath is a benchmark of 350 original, exceptionally challenging mathematics problems created by expert mathematicians (Epoch AI). Problems span number theory, analysis, algebraic geometry, and category theory. Tier 4 problems can take research mathematicians multiple days.
| Model | Tiers 1–3 | Tier 4 | Source |
|---|---|---|---|
| GPT-5.4 Pro | 50% | ~36–38% | Epoch AI |
| GPT-5.2 Pro | ~40.3% | 31% | Epoch AI |
| Gemini 3 Pro | 38% | 19% | Epoch AI |
| GPT-5.1 Thinking | ~25% | — | llm-stats |
Claude Mythos Preview is not publicly available and is only accessible through Project Glasswing, an invitation-only partner program for cybersecurity applications. Benchmarks for this model are not publicly disclosed as of May 2026.
| Benchmark | Description | Source |
|---|---|---|
| GPQA Diamond | Graduate-level science questions (PhD difficulty) | Google Research |
| MMLU-Pro | Extended multi-task language understanding (harder than MMLU) | TIGER-Lab |
| Arena Elo | Crowdsourced human preference ranking | lmarena.ai |
| HLE | Humanity's Last Exam — expert-level questions | Scale AI |
| SWE-bench Verified | Real GitHub issue resolution (human-verified subset) | SWE-bench |
| SWE-bench Pro | More challenging subset of SWE-bench | SWE-bench |
| LiveCodeBench | Live competitive programming problems (not in training data) | LiveCodeBench |
| AIME 2025 | American Invitational Mathematics Examination | MAA |
| ARC-AGI-2 | Abstract reasoning challenge (fluid intelligence) | ARC Prize |
| MMMU / MMMU-Pro | Multi-discipline multimodal understanding | MMMU |
| IFEval | Instruction-following evaluation | Google Research |
| FrontierMath | Expert-level research mathematics (Epoch AI) | Epoch AI |
AI-powered tools for software development, from IDEs and CLI tools to API providers and IDE extensions.
Integrated Development Environments with built-in AI capabilities.
| IDE | Platform | Version | Release Date | Pricing | Key Features | GitHub |
|---|---|---|---|---|---|---|
| Firebase Studio | Web | - | - | Free (3 workspaces, up to 30 with Google Developer Program) | Cloud-based, Gemini, MCP | 🔗 |
| Lingma IDE (通义灵码) | Windows, macOS | - | - | Free (download) | Built-in agent, MCP tool use, terminal command execution | ❌ |
| Tonkotsu | Windows, macOS | - | - | Free (during early access) | Team of agents, workflow | 🔗 |
| OpenCode | Windows, macOS, Linux | - | - | Free (OSS) | Terminal, desktop, IDE extension, multi-provider | 🔗 |
| Codex app | Windows | - | 2026-03-04 00:00 UTC | Included with Codex plans | Multiple agents, isolated worktrees, reviewable diffs, CLI and IDE interop | 🔗 |
| Visual Studio | Windows, macOS | 17.14.12+, 18.1.0+ | 2026-01-06 00:00 UTC | Free / $250/yr | Gemini 3 Flash integration, faster performance, zero-migration upgrades, real-time profiler agent | ❌ |
| IntelliJ IDEA | Windows, macOS, Linux | 2025.3.2 | 2026-01 | Free / $149/yr | Java 24 support, Kotlin K2 mode, performance and memory improvements | ❌ |
| IBM Bob | Cross-platform | GA (April 28, 2026) | 2026-04-28 | Free trial + Enterprise plans | Multi-model orchestration, full SDLC, 45% productivity gain | 🔗 |
| PolyAI ADK | PolyAI | GA (April 22, 2026) | 2026-04-22 | Enterprise CX | AI-native dev, Cursor/Claude Code integration | 🔗 |
| Superset | Windows, macOS, Linux | - | 2026-05-22 | Free (OSS, YC P26) | Agentic IDE for running coding agents in parallel, git worktree management, remote workspaces, task tracking | 🔗 |
| Creor | Windows, macOS, Linux | - | 2026 | Free / Paid | AI-native IDE with specialized agents (Build, Plan, Explore), agent auto-routing, MCP support | 🔗 |
| Sinew | Windows, macOS, Linux | - | 2026-05-13 | Free (MIT) | Desktop AI coding harness, Tauri 2, multi-provider, agent swarm (2-8 agents), MCP, skills | 🔗 |
| JAT | Windows, macOS, Linux | - | 2026-04-14 | Free (MIT) | Self-contained agentic IDE, 20+ parallel agents, task management, unified environment | 🔗 |
| Qoder 1.0 ⭐ | Windows, macOS, Linux | 1.0 | 2026-06-16 00:00 UTC | Free / API | Alibaba Cloud Autonomous Development Desktop; Quest window for agent command; editor + agent parallel workspaces; GA 2026-06-16 | 🔗 |
| Editor | Platform | Version | Release Date | Pricing | Key Features | GitHub |
|---|---|---|---|---|---|---|
| Zed | macOS, Windows, Linux | 0.226.3 | 2026-03-03 00:00 UTC | Free (OSS) + Copilot $10/mo | Fast, collaboration, Gemini and Claude, Zeta AI, agent thread history, edit prediction providers, self-hosted OpenAI-compatible servers | 🔗 |
| Dyad | Windows, macOS, Linux | - | - | Free (OSS) | Local generation, BYO keys | 🔗 |
| Memex | macOS, Windows | - | - | Freemium (Free + $10/mo) | Agentic, browser↔desktop | 🔗 |
| IDE | Platform | Version | Release Date | Pricing | Autonomous | MCP | GitHub |
|---|---|---|---|---|---|---|---|
| Cursor | Windows, macOS, Linux | Changelog 2026-06-30 | 2026-06-30 00:00 UTC ⭐ | Free / Pro $20/mo / Pro+ $60/mo / Ultra $200/mo; Team marketplaces now support Team MCPs and organization groups | ✅ | ✅ | ❌ |
| Windsurf | Windows, macOS, Linux | Quota plans | 2026-03-19 00:00 UTC | Free / Pro $20/mo / Teams $40/user/mo / Max $200/mo | ✅ | ✅ | ❌ |
| Trae | macOS, Windows | - | - | Free | ❌ | ❌ | 🔗 |
| PearAI | Windows, macOS, Linux | - | - | Free (OSS) | ✅ | ❌ | 🔗 |
| Void | Windows, macOS, Linux | - | - | Free (OSS) | ✅ | ✅ | 🔗 |
| Kiro | Windows, macOS, Linux | 1.0.52 | 2026-06-25 00:00 UTC ⭐ | Free preview; custom agents, permissions, agent focus mode, natural-language hooks, dockable chat tabs; auto-updates paused for 1.0.x | ✅ | ✅ | 🔗 |
| VS Code Agents | Windows, macOS, Linux | Insiders | 2026-04-21 | Free | ✅ | ✅ | 🔗 |
Cursor vs. Windsurf Pricing (June 2026): Cursor Pro is $20/month, Pro+ is $60/month, and Ultra is $200/month. Windsurf Pro is $15/month and Max is $200/month. Cursor tightens premium request quotas quarterly; heavy users often spend $40–50/mo in overages. Windsurf offers better value at the Pro tier with fewer quota restrictions. For a full alternatives analysis, see Cursor Alternatives & Better Value ↓.
Cursor Pro is $20/month, while several official pricing pages and product docs show lower-cost or BYOK alternatives with similar agent workflows, local-model support, or broader IDE coverage.
| Tool | Free Tier | Pro / Paid | Premium / Max | Notes |
|---|---|---|---|---|
| Cursor | 2K completions + 50 premium req | $20/mo (Pro) / $60/mo (Pro+) | $200/mo (Ultra) | Heavy users report $40–60/mo effective cost; quotas tightened June 2025 |
| Claude Code | Limited | $20/mo (Pro) | $100/mo (Max 5x) / $200/mo (Max 20x) | Significantly higher limits than Cursor Pro at same price |
| GitHub Copilot | Limited AI Credits | $10/mo (Pro) | $39/mo (Pro+) / $39/user (Enterprise) | Usage-based billing since June 1, 2026; AI Credits consumed per token; free models removed; no credit rollover |
| Cline (BYOK) | Free (bring API key) | Free (pay API costs ~$5–15/mo) | Unlimited | 5M+ installs; zero subscription cost |
| Windsurf | Light quota | $20/mo (Pro) / $40/user/mo (Teams) | $200/mo (Max) | Quota-based plans replaced older credit plans in March 2026 |
| Continue.dev | Free (OSS) | Free (BYOK) | Unlimited | No subscription; supports Ollama, LM Studio, all cloud APIs |
| Void | Free (OSS) | Free (BYOK) | Unlimited | VS Code fork + MCP; privacy-first, no telemetry |
| Aider | Free (BYOK) | Free (BYOK) | Unlimited | Git-native; used by serious OSS contributors |
| OpenHands | Free (OSS) | Free (self-host) | Unlimited | Full SDLC agent; MCP support |
| Rank | Tool | $/Month (Real) | SWE-bench | Limits | Verdict |
|---|---|---|---|---|---|
| 🥇 1 | GitHub Copilot Pro | $10 | 55% | Usage-based AI Credits (incl. allowance) | Best value for light-moderate users |
| 🥈 2 | Cline (BYOK) | ~$5–15 (API) | 80.8% | Unlimited | Best value for power users |
| 🥉 3 | Windsurf Pro | $15 | — | 25 credits/mo | Best value paid Pro tier |
| 4 | Claude Code Pro | $20 | 80.8% | Higher than Cursor Pro | Best autonomy at $20/mo |
| 5 | Continue.dev | $0 | — | Unlimited | Best free option |
| 6 | Cursor Pro | $20–60 (effective) | Unpublished | Caps hit easily | Good UX, poor value at scale |
| 7 | Claude Code Max | $100 | 80.8% | 5× limits | Outclasses Cursor Ultra ($200) |
This comparison uses official pricing pages, official changelogs, official extension marketplaces, and official GitHub repositories. Unofficial social feedback is not used for eligibility, pricing, or benchmark claims.
🥇 Cline (BYOK) — Best Overall Value Free VS Code extension with zero subscription cost. You pay only for your API usage (~$5–15/month for typical developer workloads). Achieved 80.8% SWE-bench Verified — identical to Claude Code. Supports 100+ providers including local Ollama/LM Studio models, meaning you can run it for $0/month with a local model like Qwen3-Coder or DeepSeek-Coder. Over 5 million installs make it the most-adopted free agent tool. Full MCP support.
🥈 Claude Code — Best Paid Alternative Anthropic's purpose-built autonomous coding CLI. At $20/mo (Pro), usage limits are substantially higher than Cursor Pro at the same price. Max 5× ($100/mo) offers more headroom than Cursor Ultra ($200/mo) at half the cost. Claude Code produces code described as ~30% less rework vs. Cursor. Won 67% of blind quality comparisons against Cursor. Integrates natively into VS Code and JetBrains via extension, plus terminal. Multi-session support, computer use (research preview), sub-agents.
🥉 GitHub Copilot — Best for Teams & IDE-Native Work At $10/month (Pro), GitHub Copilot requires no fork, no new editor, no learning curve. Works inside existing VS Code, JetBrains, Vim, Neovim, and more. Since June 1, 2026, billing switched to usage-based AI Credits; free models are no longer included, and unused credits do not roll over. Pro+ at $39/mo adds enterprise policy controls. Best for developers who want AI assistance without changing their workflow. Dominates for boilerplate and quick completions.
Continue.dev — Best Free Fully-Featured Option 100% open-source, works in VS Code and JetBrains. BYOK means no subscription cost ever. Supports Ollama, LM Studio, and all major cloud APIs. Chat, autocomplete, and agent mode. Active development community. No telemetry by default.
Void — Best Privacy-First Fork Open-source VS Code fork with full MCP support, BYOK, and no telemetry. Feature-comparable to Cursor with zero subscription cost. Best for teams that need data sovereignty or work in air-gapped environments.
🎯 Bottom Line for Sharif's Use Case: If you use LM Studio and prefer local models, Cline or Continue.dev (both BYOK, both free) give you unlimited usage at zero subscription cost. For cloud-based work, Claude Code Pro ($20) outclasses Cursor Pro ($20) with higher limits and better autonomous task completion. GitHub Copilot ($10) is the obvious choice if budget is the primary driver.
| IDE | Platform | Version | Release Date | Pricing | Self-Hostable | Best For | GitHub |
|---|---|---|---|---|---|---|---|
| Replit 3 | Web | - | - | Free Starter, Core $20/mo, Pro $100/mo | ❌ | Learning/Prototyping | ❌ |
| Bolt.new | Web | - | - | Free, Pro $20-25/mo, Teams $30/user/mo | ❌ | Quick apps | ❌ |
| Bolt.diy | Self-hosted | - | - | Free (MIT), bring your own API | ✅ | Self-hosted | 🔗 |
| Lovable | Web | - | - | Free (5 credits/day), Pro $25/mo, Business $50/mo | ❌ | UI/Full-stack | ❌ |
| v0 | Web | - | - | Free ($5 credits/mo), Premium $20/mo, Teams $30/user | ❌ | React components | ❌ |
| Gitpod | Web | - | - | Free + Paid | ❌ | Cloud dev environments | ❌ |
| Rork | Web | - | - | Free & Paid (credits) | ❌ | Mobile apps (iOS/Android) | ❌ |
| Google Stitch | Web | - | 2026-03 | Free (Google account, 550 gen/mo) | ❌ | UI design, Figma/React export | ❌ |
| Google Antigravity 2.0 | Windows, macOS, Linux (CLI), Web (IDE) | 2.0 | 2026-05-19 00:00 UTC | Google AI Pro $19.99/mo / Ultra $99.99/mo (5×) or $200/mo (20×) | Agent-first platform: desktop app + CLI (agy) + SDK + Managed Agents API; multi-agent orchestration, parallel tasks, voice control |
🔗 | |
| Jules | Web | - | 2025-05-20 00:00 UTC | Free beta, higher limits on Google AI Pro / Ultra | Async repo agent, reviewable diffs, GitHub integration | ❌ |
Command-line AI tools for autonomous coding and terminal enhancement.
| Tool | Platform | Pricing | Key Features | GitHub |
|---|---|---|---|---|
| Aider | Windows, macOS, Linux | Free | Gold standard, Architect mode, thinking tokens | 🔗 |
| Claude Code 2.2.1+ | macOS, Linux, Windows | Free + API | Sonnet 5 default support via Claude Code 2.1.197+; Opus 4.8 support; Fable 5 access restored globally 2026-07-01 00:00 UTC; simple mode file editing, multi-session support, /autofix-pr, plan mode, computer use (research preview), CLAUDE.md skills & sub-agents |
🔗 |
| Codex CLI ⭐ | Windows, macOS, Linux | Included with ChatGPT Plus/Pro/Business/Edu/Enterprise or API key | Rust-native local coding agent; sandbox, approval modes, subagents, command autocomplete, workspace search; v0.142.5 release 2026-07-01 patched Responses WebSocket trace logging | 🔗 |
| Junie CLI ⭐ | Windows, macOS, Linux | Free (BYOK) | LLM-agnostic, JetBrains IDE integration (ACP), GA 2026-06-17, agentic debugging, PR review, local model support (LiteLLM/LMStudio/Ollama) | 🔗 |
| Goose | Windows, macOS, Linux | Free (Apache-2.0) | MCP, extensible, desktop app, 25+ providers | 🔗 |
| Grok Build | Windows, macOS, Linux | $300/mo (SuperGrok/X Premium Plus) | Plan mode, Arena mode, 2M context window, security-focused | ❌ |
| GPT-Pilot | Windows, macOS, Linux | Free | Full dev team simulation | 🔗 |
| OpenHands | Windows, macOS, Linux | Free | Cloud agents, MCP | 🔗 |
| Mentat | Windows, macOS, Linux | Free | Multi-file coordination | 🔗 |
| SERA | Linux, macOS | Free (Apache 2.0) | Open-source coding agent, 200K synthetic trajectories | 🔗 |
| Kimi K2.7 Code ⭐ | Windows, macOS, Linux | $0.95 / $4.00 per 1M | Coding-focused agent; 30% fewer thinking tokens vs K2.6; Modified MIT; 262K context; released 2026-06-12 | 🔗 |
| Kilo for GitHub ⭐ | GitHub (Cloud) | Free (Kilo credits) | @kilocode-bot coding agent in GitHub issues/PRs; Cloud Agent backend; reads full thread context; released 2026-06-18 | 🔗 |
| Hermes Agent v0.17.0 ⭐ | Windows, macOS, Linux | Free (OSS) | Major desktop app update; live subagent watch-windows; VS Code theme support; 199K GitHub stars; released 2026-06-19 | 🔗 |
| Kimi Code CLI | Windows, macOS, Linux | Free (MIT) | TypeScript CLI by Moonshot AI; successor to kimi-cli; powered by Kimi K2.6 (1T MoE); subagent support (coder/explore/plan); 262K context; SWE-bench Pro 58.6%; ACP (Zed, JetBrains); released 2026-06-06 | 🔗 |
| Vibe (Mistral) | Windows, macOS, Linux | Free (Le Chat) / API | Agentic coding + work agent, multi-model Mistral harness, VS Code extension, Code Mode web surface, parallel sessions | 🔗 |
| Google Agents CLI ⭐ | Windows, macOS, Linux | Free | Unified ADLC for Google Cloud Agent Platform; scaffold/run/evaluate/deploy agents; released 2026-04-22 | 🔗 |
| Peri ⭐ | Windows, macOS, Linux | Free (Apache-2.0) | Rust terminal coding agent; Claude Code compatible; multi-LLM; MCP & ACP support; 95%+ prompt cache; released 2026-03-20 | 🔗 |
| AI Dev Kit | Cross-platform | Free | 59 skills, 33 agents, TDD, security audit, CI/CD | 🔗 |
| Tool | Developer | Pricing | Best For | GitHub |
|---|---|---|---|---|
| Gemini CLI |
Free (now shut down) | Google ecosystem & Gemini models in-terminal; stopped serving requests June 18, 2026 for AI Pro/Ultra/free users; migrate to Antigravity CLI | 🔗 | |
Antigravity CLI (agy) ⭐ |
Free (~20 req/day) / Pro/Ultra higher limits | Agentic coding, multi-agent orchestration, Go-based replacement for Gemini CLI; released 2026-06-18 | 🔗 | |
| Google Colab CLI ⭐ | Free (Colab account) | Run Python on remote Colab GPUs/TPUs from local terminal; AI agent compatible; released 2026-06-06 | 🔗 |
⚠️ Anthropic June 15 Billing Split: As of 2026-06-15, Anthropic moved Claude Agent SDK,claude -p, and Claude Code GitHub Actions to separate metered credits billed at full API rates. Pro subscribers get $20/mo credit; Max 20x gets $200/mo. Interactive Claude Code and Claude.ai remain on standard subscription. Fable 5 was included on Pro/Max/Team/Enterprise subscription plans at no extra cost through June 22; after that date, usage credits are required.
| Cursor CLI | Cursor | Free tier | Terminal + IDE bridge | 🔗 | | Qwen Code | Alibaba | Free | Qwen optimization | 🔗 | | Qodo CLI | Qodo | Free tier | Testing, review & agent workflows | 🔗 |
Gemini CLI (deprecation & security notes): As of June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions have stopped serving requests for Google AI Pro, Ultra, and free Gemini Code Assist for individuals, per Google's official announcement. Google ecosystem users should migrate agentic workflows to Antigravity CLI (
agy). Enterprise Code Assist Standard/Enterprise licensees retain access. Current CLI generations add offline search, bundled ripgrep, optional color-accessible themes, interactive shell tool invocation, and an A2A-style agent registry surface in the tool (see Gemini CLI changelogs). Treat upgrades through v0.39.1+ as mandatory maintenance: public advisories described a critical remote-code path affecting earlier builds (widely reported as CVSS 10.0); confirm your installed build against the release notes for your package channel. Source: Google Developers Blog
AI coding CLI tools categorized by their primary language support. All tools below accept plain English prompts.
| Tool | Primary Languages | Multi-Language | Local LLM | Cloud API | Pricing | GitHub |
|---|---|---|---|---|---|---|
| Aider | Python, JS, TS, Go, Rust, Ruby, Java, C/C++ | ✅ (100+ langs) | ✅ (Ollama, LM Studio) | ✅ | Free (OSS) | 🔗 |
| Claude Code | All (polyglot) | ✅ | ❌ | Claude API ($3–$15/M) | Free tool / API cost | ❌ |
| Codex CLI | Python, JS, TS, Bash | ✅ | ❌ | OpenAI API | Free OSS / API cost | 🔗 |
| OpenHands | Python, JS, TS, Go, Rust, Java | ✅ | ✅ | ✅ | Free (OSS) | 🔗 |
| Goose (Block) | Polyglot (25+ providers) | ✅ | ✅ (Ollama, LM Studio) | ✅ | Free (OSS) | 🔗 |
| Continue | Polyglot (VS Code, JetBrains) | ✅ | ✅ (Ollama, LM Studio) | ✅ | Free (OSS) | 🔗 |
| Qwen Code | Python, JS, TS, Go, Java | ✅ | ✅ (Qwen models) | ✅ | Free (OSS) | 🔗 |
| Devstral CLI | Python, JS, TS, Go, Rust | ✅ | ✅ | Mistral API | Free OSS model / API cost | ❌ |
| OpenCode | Polyglot | ✅ | ✅ | ✅ | Free (OSS) | 🔗 |
| Mentat | Python, JS, TS, Go | ✅ | ❌ | OpenAI API | Free (OSS) | 🔗 |
| Amp (Sourcegraph) | All | ✅ | ❌ | ✅ | Free / Enterprise | ❌ |
Purpose-built CLI tools for coding across specific languages or polyglot multi-stack workflows.
| Tool | Language Focus | Platform | Pricing | Key Features | GitHub |
|---|---|---|---|---|---|
| Aider | Polyglot (Python, JS, TS, Go, Rust, any) | All | Free (BYOK) | Git-native multi-file edits, Architect mode, repo maps, thinking tokens | 🔗 |
| Claude Code | Polyglot | All | Free + API | Computer use, sub-agents, CLAUDE.md skills, Opus 4.7, multi-session | 🔗 |
| Codex CLI | Python, JS, TS | All | Free (OpenAI account) | Sandbox execution, approval modes, OpenAI models | 🔗 |
| OpenHands | Python, JS, TS, Go, Rust | All | Free (OSS) | Full SDLC agent, MCP, local LLM via Ollama | 🔗 |
| Goose | Polyglot | All | Free (Apache-2.0) | 25+ providers, MCP, extensible extensions, desktop app | 🔗 |
| Continue | Polyglot | All | Free (OSS) | VS Code + JetBrains, custom models via Ollama/LM Studio | 🔗 |
| Qwen Code | Python, JS, TS, Go | All | Free | Optimized for Qwen3-Coder 480B, Apache 2.0 | ❌ |
| Mentat | Polyglot | All | Free | Multi-file coordination, context-aware diffs | 🔗 |
| AI Dev Kit | Polyglot | All | Free | 59 skills, 33 agents, TDD, security audit, CI/CD pipeline | 🔗 |
| Devstral CLI | Python, JS, TS, Go | All | Free (Mistral free tier) | Mistral's open coding model, OpenRouter free access | ❌ |
| Junie CLI | Polyglot | All | Free (BYOK) | LLM-agnostic, JetBrains IDE integration, MCP | 🔗 |
| SERA | Python, JS, TS | Linux, macOS | Free (Apache 2.0) | Open-source coding agent, 200K synthetic trajectories | 🔗 |
| Tool | Platform | Pricing | Key Features |
|---|---|---|---|
| Warp Terminal | macOS, Linux, Windows | Free | AI Agents, workflow sharing, Warp Drive automation |
| Microsoft Intelligent Terminal | Windows | Free (MIT) | AI agent pane (Agent Client Protocol), automatic error detection, GitHub Copilot CLI default, Windows Terminal fork — 2026-06-02 ⭐ |
| Starship | Cross-platform | Free (OSS) | Fast cross-shell prompt; rich plugins & AI-adjacent ecosystem |
| Fig | macOS, Linux | Free | Autocomplete, AI suggestions |
| Google Colab CLI ⭐ | Cross-platform | Free (Colab account) | Run Python on remote Colab GPUs/TPUs from local terminal; AI agent compatible; released 2026-06-06 |
Extensions and plugins that add AI capabilities to existing IDEs.
| Add-on | Platform | Pricing | Context | Best For | GitHub |
|---|---|---|---|---|---|
| GitHub Copilot ⭐ | VS Code, JetBrains, Vim | Usage-based AI Credits since June 1, 2026 | Large | General coding; Kimi K2.7 Code available in model picker (July 1, 2026); Copilot Vision GA (July 1, 2026); enterprise session streaming/API, AI credit pools, and Copilot-only Gemini 2.5 Pro / Gemini 3 Flash deprecation on 2026-07-31 | 🔗 |
| Supermaven | VS Code, JetBrains, Neovim | Free / $10/mo | 1M | Large codebases | ❌ |
| Codeium | VS Code, JetBrains, Vim | Free / $15/mo / $60/mo | Medium | Free alternative | ❌ |
| Continue | VS Code, JetBrains | Free (OSS) | Custom | Self-hosted | 🔗 |
| Cody | VS Code, JetBrains, Web | Free (discontinued) / Enterprise Starter $19/mo / Enterprise $59/mo | Enterprise | Code search | 🔗 |
| Tabnine | VS Code, JetBrains, VS, Eclipse | Free / $39/mo | Local | Privacy | ❌ |
| Tabby | VS Code, JetBrains, Vim, Neovim | Free (OSS) | Self-hosted | Self-hosted code completion | 🔗 |
| Add-on | Pricing | Autonomous | MCP | Best For | GitHub |
|---|---|---|---|---|---|
| Codex | Free (with ChatGPT Plus $20/mo or Pro $200/mo) | ✅ | ✅ | OpenAI's official coding agent | 🔗 |
| Cline | Free | ✅ | ✅ | Full agent | 🔗 |
| GitHub Copilot (Agent Mode) ⭐ | Usage-based AI Credits since June 1, 2026 | ❌ | Guided agent workflows; auto-opens PRs with self-review; Kimi K2.7 Code available; Copilot CLI in GitHub Actions now supports GITHUB_TOKEN with copilot-requests: write |
🔗 | |
| RooCode | Free/Pro | ❌ | Complex tasks | 🔗 | |
| JetBrains AI Assistant | $10/mo (Pro) | ✅ | ❌ | JetBrains-quality AI now in VS Code (public preview); multi-file edits, Mellum LLM | ❌ |
| ReSharper for VS Code | Included with JetBrains subscription | ❌ | ❌ | C# code analysis + refactoring inside VS Code/Cursor; released 2026-03-05 | ❌ |
| Keploy | OSS/Enterprise | ❌ | ❌ | Testing | ❌ |
| Microsoft Foundry Toolkit for VS Code ⭐ | Free (Azure account) | ✅ | ✅ | End-to-end Hosted Agent lifecycle; GA 2026-04-16; consolidated AI Toolkit + Foundry extension; LangGraph samples; released 2026-06-02 | 🔗 |
| VS Code ACP Client ⭐ | Free (MIT) | ✅ | ✅ | Agent Client Protocol client; connects 11+ AI agents (Claude Code, Codex, Cursor, etc.) to VS Code; released 2026-06-10 | 🔗 |
| Prompt Foundry ⭐ | Free | ❌ | ✅ | VS Code/Cursor extension; modular prompt blocks with Liquid syntax; MCP server for context updates; released 2026-06-18 | 🔗 |
| GrackerAI ⭐ | Free | ❌ | ❌ | Validates llms.txt, robots.txt, AI visibility signals in VS Code; released 2026-06-16 | ❌ |
| A11yResolver ⭐ | Free (beta) | ❌ | ❌ | VS Code extension for accessibility remediation; AI agent flags WCAG issues; released 2026-06-18 | ❌ |
Claude Tag (Anthropic, June 2026): Not an IDE extension but worth noting here. Claude Tag brings an always-on AI teammate to Slack. Each channel gets an isolated Claude identity with persistent memory. Available on Team and Enterprise plans. Source: 🔗
| Add-on | Pricing | Claude Agent | Junie (GA) | Best For |
|---|---|---|---|---|
| JetBrains AI Assistant | $10/mo (Pro), $249/yr (Ultimate) | ✅ | ✅ | Deep IDE integration; Junie GA 2026-06-17 |
| JetBrains Claude Agent | Included in subscription | ✅ | — | Native agent |
| JetBrains Junie ⭐ | Included in subscription | — | ✅ | GA 2026-06-17; ACP-based IDE integration; agentic debugging; BYOK; local model support; IntelliJ/WebStorm/PyCharm/GoLand/CLion |
JetBrains Marketplace Security (June 2026): 15 malicious AI plugins removed from JetBrains Marketplace; 7 publisher accounts terminated; remote kill-switch triggered. If you installed any affected plugins before June 17, 2026, revoke exposed API keys immediately.
Services for accessing AI models via API. This pass prioritizes the cheapest verified public text APIs and the free or trial access that is explicitly documented by each provider.
| Provider | Cheapest verified public text model | Direct price | Free or trial access | Notes | Source |
|---|---|---|---|---|---|
| Groq | Llama 3.1 8B Instant | $0.05 input / $0.08 output | GroqCloud Free plan ($0) | Lowest verified public direct text API price in this pass | 🔗, 🔗 |
| Google Gemini | Gemini 2.5 Flash-Lite | Free tier: free of charge; paid tier $0.10 / $0.40 | Gemini Developer API free tier | Batch and Flex drop to $0.05 / $0.20 | 🔗 |
| DeepSeek | DeepSeek-V4-Flash | $0.14 input / $0.28 output | No standing free tier listed on pricing page | Cache-hit input is $0.0028 | 🔗 |
| Mistral AI | Mistral Small 4 | $0.15 input / $0.60 output | Free mode enabled by default, no credit card required | Good low-cost coding and multimodal option | 🔗, 🔗 |
| Cohere | Command R | $0.15 input / $0.60 output | Trial API keys are free but limited; Command A+ is free until rate limits are reached | Low-cost RAG and tool use option | 🔗, 🔗, 🔗 |
| OpenAI | GPT-5.4 nano | $0.20 input / $1.25 output | OpenAI quickstart includes one free test API request | Batch and Flex drop to $0.10 / $0.625 | 🔗, 🔗 |
| Anthropic | Claude Haiku 4.5 | $1.00 input / $5.00 output | New users receive a small amount of free credits | Lowest-cost Claude API model | 🔗, 🔗 |
| xAI | Grok 4.3 | $1.25 input / $2.50 output | No public free tier listed on pricing page | Grok 4.5 is $2.00 / $6.00 | 🔗 |
| Provider | Models | Key Features |
|---|---|---|
| OpenRouter | 400+ | Unified API, crypto/fiat, model rankings, fallback routing, Speech & Transcription APIs, Model Fusion (launched 2026-06-12), private models, enterprise workspace controls; new models: Claude Opus 4.8, Step 3.7 Flash, MiniMax M3, Qwen3.7-Plus, NVIDIA Nemotron 3 Ultra, Microsoft MAI models, Nano Banana 2, Nano Banana Pro, GLM 5.2, Cohere North Mini Code (free), Nex-N2-Pro (free), GPT-Image-1.5, GPT-Image-2, Poolside Laguna, Qwen3-Coder (June 2026) |
| Hugging Face | Thousands | Serverless inference, free tier |
| LiteLLM | 100+ | Open-source proxy; single OpenAI-compatible API for any provider |
| Portkey | 200+ | Gateway: load balance, fallbacks, caching, observability |
Services that aggregate multiple models through a unified API, often with load balancing, caching, and cost optimization features.
| Provider | Models | Pricing | Key Features |
|---|---|---|---|
| Together AI | 200+ | $0.10-$2.50/1M | Broadest open-weight catalog (text/image/video/audio/embeddings), NVIDIA Blackwell-optimized |
| Fireworks AI | 50+ | $0.10-$1.50/1M | Production-grade open-source inference, 90% cheaper than closed providers, NVIDIA Blackwell |
| Groq | 20+ | $0.05-$0.90/1M | Fastest LPU inference: 476 tok/s on 120B, <100ms TTFT, deterministic latency |
| Cerebras | 10+ | $0.10-$1.00/1M | Wafer-scale engine: ~3,000 tok/s, 5× faster than NVIDIA Blackwell, 80–150ms TTFT |
| NVIDIA NIM | 91+ | Free (cloud endpoints) / self-host | Broadest model variety: LLMs + vision + audio + bio + climate; self-hostable Docker containers |
| Vercel AI Gateway ⭐ | 100+ | $5/mo free credits; paid credits at provider list rates, zero markup | Unified API for text, image, video, realtime, speech, embeddings, and reranking; BYOK on paid tier; model fallbacks, provider routing, budgets, observability, and Routing Rules beta |
| Mercury 2 (Inception Labs) ⭐ | 1 | $0.25/$0.75/1M | Diffusion LLM; 1,009 tok/s on Blackwell; tunable reasoning; 128K context; released 2026-05-12 |
| Anyscale | 100+ | $0.20-$2.00/1M | OpenAI-compatible, HIPAA/SOC 2/EU data residency, enterprise contracts |
| Replicate | 50+ | $0.20-$2.00/1M | Model-as-a-service, versioning, GPU rental |
| OctoAI | 20+ | $0.25-$2.00/1M | High-performance inference |
| Predibase | 12+ | $0.30-$2.50/1M | Fine-tuning + multi-LoRA serving |
| RunPod Serverless | 10+ | $0.20-$2.00/1M | GPU serverless, flexible pricing |
| CoreWeave | 8+ | Custom | Enterprise-grade, dedicated GPUs |
| Infron ⭐ | 100+ | Up to 35% off direct pricing | Enterprise-grade unified API; provisioned throughput plan; broad model selection; released 2026-06-10 |
| Aethir Mesh ⭐ | 10+ open-source | Pay-per-use | Decentralized GPU infrastructure; single API for open-source models (DeepSeek V4, Kimi K2.6, GLM-5.1, etc.); released 2026-06-12 |
| Provider | Models | Pricing | Best For |
|---|---|---|---|
| DeepInfra | 50+ | Free tier + $0.10-1.00/1M | Free tier, diverse models |
| TensorZero | 30+ | Free tier + $0.15-1.50/1M | Cost optimization, caching |
| Scale AI | 25+ | Custom pricing | Enterprise, managed services |
| AssemblyAI | 15+ | Free tier + pay-per-use | Speech + text models |
| Provider | Type | A100 80GB ($/hr) | H100 80GB ($/hr) | Best For |
|---|---|---|---|---|
| Vast.ai | GPU Marketplace | ~$0.09–$0.59 | ~$0.50–$1.50 | Cheapest market pricing, diverse GPU options |
| RunPod | GPU Rental | $0.34–$0.69 (RTX 4090) / $0.60–$1.19 (A100) | $1.99–$3.29 | Flexibility, per-second billing, cost-effective fine-tuning & inference |
| Lambda Labs | Cloud GPU | $1.29–$2.79 | $2.49–$3.99 | Reliable on-demand, academic discounts, 1-Click Clusters |
| DigitalOcean | GPU Droplets | $1.29 | $2.99 | Simple fine-tuning workflows |
| Vultr | Global Cloud | $1.29–$2.80 | $1.99–$2.99 | Hourly GPU instances, global regions |
| Hyperbolic | Decentralized | ~$1.50 | ~$3.50 | Crypto/Fiat payments |
| Cerebrium | Serverless GPU | ~$2.10 | ~$3.40 | Python-native ML inference & fine-tuning |
| Modal Labs | Serverless GPU | ~$2.50–$3.40 | ~$3.95 | Fine-tuning with LoRA, distributed training; per-second billing |
| Replicate | Model-as-a-Service | $5.04 | $5.49 | Quick deployment, serverless inference |
| Together AI | AI-Native Cloud | — | — | Fast inference & fine-tuning for open models |
| Fireworks AI | Inference & Fine-tuning | — | — | Fast inference, RFT for model shaping |
| CoreWeave | Enterprise GPU Cloud | ~$2.70 | ~$6.16 | Enterprise-grade, dedicated GPUs, Kubernetes-native |
| Databricks Mosaic AI | Integrated ML Platform | — | — | Enterprise fine-tuning, governed serving, RAG |
| NVIDIA DGX Cloud | Managed AI Training | Custom | Custom | Co-engineered clusters, maximum ROI for training |
GPU pricing as of June 2026. RunPod/Vast.ai prices vary by community vs. secure cloud and spot vs. on-demand. RTX 4090 pricing from RunPod starts at $0.34/hr. Vast.ai A100 40GB from $0.52/hr. Modal Labs production workloads cost 3.75x base rates (3x non-preemptible multiplier x 1.25x US regional multiplier). CoreWeave A100/H100 pricing normalized from 8-GPU nodes. Lambda Labs A100 40GB at $1.29/hr, A100 80GB at $2.49-$2.79/hr, H100 SXM at $3.99/hr (8x) or $4.29/hr (1x). Vast.ai prices are market-set medians.
| Provider | Specialization | Speed |
|---|---|---|
| Together AI | Llama/Qwen/Mistral | Fast |
| Fireworks AI | FireAttention | Low-latency, 6 free models |
| Groq | LPU | >500 T/s |
| Cerebras | Wafer-Scale | >2000 T/s |
| NVIDIA NIM | 91 free endpoints, DGX Cloud | 20× faster than NVIDIA GPU |
Services that provide hosted access to open source models with API endpoints, offering easier deployment than self-hosting.
| Provider | Models | Pricing | Best For |
|---|---|---|---|
| Hugging Face Inference | 1000+ | Free tier + $0.10-1.00/1M | Largest selection, easy deployment |
| Replicate | 50+ | Free tier + $0.20-2.00/1M | Model-as-a-service, versioning |
| Together AI | 30+ | $0.20-3.00/1M | Optimized for open models |
| Fireworks AI | 25+ | Free tier + $0.10-1.50/1M | Low-latency, 6 free models |
| Anyscale | 100+ | $0.20-2.00/1M | Distributed inference, scaling |
| OctoAI | 20+ | $0.25-2.00/1M | High-performance GPU inference |
| Banana.dev | 15+ | $0.15-1.80/1M | Fast deployment, auto-scaling |
| Predibase | 12+ | $0.30-2.50/1M | Fine-tuning + inference |
| RunPod Serverless | 10+ | $0.20-2.00/1M | GPU serverless, flexible pricing |
| CoreWeave | 8+ | Custom pricing | Enterprise-grade, dedicated GPUs |
| Provider | Official free access | What happens after free or trial | Best For | Source |
|---|---|---|---|---|
| Google Gemini | Gemini Developer API free tier is free of charge | Gemini 2.5 Flash-Lite paid tier starts at $0.10 / $0.40 | Best official always-on free multimodal API | 🔗 |
| Groq | GroqCloud Free plan is $0 for build and test use | Developer plan is pay-per-token; Llama 3.1 8B Instant starts at $0.05 / $0.08 | Fast low-cost iteration | 🔗, 🔗 |
| Cohere | Trial API keys are free but limited; Command A+ is free until rate limits are reached | Command R starts at $0.15 / $0.60 | RAG, tool use, multilingual testing | 🔗, 🔗, 🔗 |
| Mistral AI | Free mode is enabled by default with no credit card required | Mistral Small 4 starts at $0.15 / $0.60 on Scale | Cheap coding and multimodal experiments | 🔗, 🔗 |
| Anthropic | New users receive a small amount of free credits | Haiku 4.5 starts at $1.00 / $5.00 | Testing Claude API behavior | 🔗, 🔗 |
| Vercel AI Gateway ⭐ | $5/month included on the free tier for eligible models | Paid credits use provider list rates with zero markup | Multi-provider testing behind one endpoint | 🔗 |
| OpenAI | Developer quickstart includes one free test API request | Standard API pricing starts with GPT-5.4 nano at $0.20 / $1.25 | SDK evaluation and first-request testing | 🔗, 🔗 |
AI-powered tools for automating browser and desktop tasks.
Tools and frameworks for AI-powered browser automation.
| Browser | Platform | Pricing | Open Source | Local AI | Agent/Computer Use | API Access | Multi-Agent | Parallel Sessions | Best For | GitHub |
|---|---|---|---|---|---|---|---|---|---|---|
| Norton Neo | Windows, macOS, iOS, Android | Free | ❌ | ❌ | ❌ | ❌ | ❌ | AI-native browser, Magic Box, Peek & Summarize, Smart Tab Groups | ❌ | |
| Deta Surf | macOS, Linux | Free (alpha) | ✅ | ✅ | ❌ | ✅ | ✅ | Browser + file manager + AI notebook, Surf Memory | ❌ | |
| Firefox (AI Controls) | Windows, macOS, Linux, iOS, Android | Free | ✅ | ❌ | ❌ | ❌ | ❌ | AI Controls dashboard (v148+), ChatGPT/Claude/Mistral sidebars | ❌ | |
| Dendrite | Windows, macOS, Linux | Free (OSS) | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | Developer AI browser client for agents | ❌ |
| Perplexity Comet | Windows, macOS, iOS, Android | Free / Pro $20/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Research + background tasks, voice mode, Computer Max agent | ❌ |
| ChatGPT Agent Mode | Web, iOS, Android | Plus $20/mo, Pro $200/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Full computer use: browse, code, fill forms, book travel | ❌ |
| Dia | macOS (M1+ / macOS 14+) | Free / Pro $20/mo | ❌ | ❌ | ❌ | ❌ | ❌ | Tab intelligence, Skills, browsing history AI context | ❌ | |
| Google Chrome (Auto Browse 2) | Windows, macOS, Linux, ChromeOS | Free / Gemini Pro $19.99/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Gemini 3 built-in, auto browse 2 agentic tasks, Chrome Skills saveable workflows, Universal Commerce Protocol, WebMCP support | ❌ |
| Microsoft Edge (Copilot Agent) | Windows, macOS, iOS, Android | Free / Copilot Pro $20/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Cross-tab context, voice commands, form automation, bookings | ❌ |
| Genspark | Windows, macOS, Web, iOS, Android | Free / Plus $25/mo / Pro $249/mo | ❌ | ✅ (169 local models) | ✅ | ❌ | ✅ | ✅ | Super Agent, AI slides, AI websites, deep research, Call For Me, Genspark Claw desktop app | ❌ |
| Brave Leo (AI Browser) | Windows, macOS, Linux, iOS, Android | Free / Premium $14.99/mo | ✅ (Chromium) | ✅ (Leo local) | ❌ | ❌ | ❌ | Privacy-first, zero-log AI, Skills, Memories, local models | ❌ | |
| SigmaOS (Airis) | macOS | Free / Pro (subscription) | ❌ | ❌ | ❌ | ❌ | ❌ | NL commands: "Book Airbnb in Iceland", cross-tab AI, YC-backed | ❌ | |
| Opera Neon | Windows, macOS | $19.90/mo | ❌ | ❌ | ✅ | ✅ MCP Connector (Mar 2026) | ❌ | ❌ | Agentic browsing, Aria assistant, MCP API for Claude/ChatGPT/n8n integration | ❌ |
| Opera One (Aria) | Windows, macOS, Linux, iOS, Android | Free | ❌ | ❌ | ❌ | ❌ | ❌ | Built-in Aria AI assistant, sidebar AI tools | ❌ | |
| Firefox (AI Sidebar) | Windows, macOS, Linux, iOS, Android | Free | ✅ | ❌ | ❌ | ❌ | ❌ | AI Controls dashboard (v148+), ChatGPT/Claude/Mistral sidebars | ❌ | |
| BrowserOS | Linux, macOS | Free | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | Privacy-focused, built-in MCP, agentic | 🔗 |
| Manus AI | Web (Cloud) | Free 300 credits/day / Plus $20/mo / Pro $200/mo | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | Cloud agent, full computer: code, deploy, files, search | ❌ |
| Sigma AI Browser | Windows, macOS, Linux | Free / Pro $29/mo | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | Built-in local AI agent, offline, no tracking | 🔗 |
| Fellou | Windows, macOS | Free 4 tasks/day / Pro $20/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Complex multi-step automation, agentic tasks | 🔗 |
| Arc Max | macOS, Windows | Free | ❌ | ❌ | ❌ | ❌ | ❌ | AI-enhanced browsing, pinch-to-summarize, Ask on Page | ❌ | |
| Maxthon | Windows, macOS, iOS, Android | Free / Premium | ❌ | ❌ | ❌ | ❌ | ❌ | MaxAsk AI answers, built-in VPN, ad-blocker, resource sniffer | ❌ | |
| ChatGPT Atlas | macOS | Free (with ChatGPT subscription) | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | OpenAI integration, macOS computer use overlay | 🔗 |
| AnythingLLM | Windows, macOS, Linux | Free (OSS) | ✅ | ✅ | ✅ (local API) | ❌ | ❌ | All-in-one desktop AI, document chat, local + API | 🔗 | |
| BrowserGPT | iOS, Android | Free / Premium | ❌ | ❌ | ❌ | ❌ | ❌ | Mobile-first AI browser | ❌ | |
| Sidekick Browser | Windows, macOS, Linux | Free / Pro $10/mo | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | AI assistant, natural language tab management, summarize, automate tasks | ❌ |
| Browser Operator | Windows, macOS | Free (OSS) | ✅ | ✅ (Ollama, 100+ models) | ✅ | ❌ | ✅ | ✅ | Multi-agent automation, privacy-first local processing, persistent memory | 🔗 |
| Browserless Agent ⭐ | Cloud (MCP) | Free tier / paid | ❌ | ❌ | ✅ | ✅ (MCP-native) | ❌ | ❌ | Fastest MCP browser agent; stateful sessions; command batching; released 2026-06-12 | 🔗 |
| Kane CLI ⭐ | Windows, macOS, Linux | Free | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | Browser automation testing for AI agents; deterministic pass/fail; released 2026-05-24 | 🔗 |
| Browsewright ⭐ | Python | Free (MIT) | ✅ | ✅ | ✅ | ✅ (MCP) | ❌ | ❌ | Open-source browser agent; LLM-driven form filling; structured JSON output; released 2026-06-16 | 🔗 |
| Lightpanda Agent ⭐ | Windows, macOS, Linux | Free (OSS) | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | Native headless browser agent; PandaScript reproducible scripts; LLM at build-time not runtime; released 2026-06-17 | 🔗 |
| Veilbrowser ⭐ | TypeScript | Free (MIT) | ✅ | ✅ | ✅ | ✅ (MCP) | ❌ | ❌ | Stealth browser for AI agents; real Chrome over raw CDP; passes sannysoft 57/57; released 2026-06-09 | 🔗 |
| Dev Browser ⭐ | Claude Code plugin | Free | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | Browser automation plugin for Claude Code; stateful Playwright wrapper; released 2026-06-19 | 🔗 |
| Intuned ⭐ | Python/Cloud | Free tier / paid | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | YC-backed browser automation as code; AI agent generates Playwright code; auto-healing; released 2026-06-08 | ❌ |
| Webwright (Microsoft) ⭐ | Python | Free (MIT) | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | SOTA on long-horizon web tasks; terminal-based; Codex/Claude Code plugins; released 2026-04-08 | 🔗 |
| SimuLang ⭐ | Python/TypeScript | Free (OSS) | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | Playwright for the entire desktop; browser + native apps + OS workflows; released 2026-05-19 | 🔗 |
| Extension | Pricing | Free | Multi-Agent | Best For | GitHub |
|---|---|---|---|---|---|
| OpenDia | Free (OSS) | ✅ | ❌ | Open alternative to Dia/Comet; anti-detection for X/LinkedIn/Facebook; works with Chrome/Firefox/Edge/Brave | 🔗 |
| Monica.im | Freemium (Free + ~$9/mo) | ❌ | ✅ | Chrome extension, no browser switch | ❌ |
| Harpa AI | Free | ✅ | ❌ | Automation recipes | 🔗 |
| MultiOn | Free/Paid | ✅ | Complex tasks | 🔗 | |
| NanoBrowser | Free | ✅ | ✅ | Local control, Ollama | 🔗 |
| Neobrowser | Free (OSS) | ✅ | ❌ | Local LLMs via Ollama, privacy-first, Chrome/Edge | ❌ |
| Open Operator | Free | ✅ | ❌ | Browserbase-powered, open NL browser control | 🔗 |
| Openator | Free (OSS) | ✅ | ❌ | Docker-based headless NL browser agent | 🔗 |
Consumer-focused AI browsers (for example Perplexity Comet) typically bundle chat with browsing but enforce vendor-side quotas on usage tiers. The stacks below target developers: scale via your LLM keys (BYOK where offered), metered cloud APIs, or self-hosted / OSS runtimes so automation limits follow your infrastructure and providers—not the same caps as end-user browser apps.
| Library | Language | Pricing | Best For | API Access | Multi-Agent | Parallel Sessions | GitHub |
|---|---|---|---|---|---|---|---|
| Chrome DevTools MCP | TypeScript | Free (OSS) | AI web debugging, 29 DevTools | ❌ | ❌ | ❌ | 🔗 |
| Cloudflare Browser Run | Cloud API | Free Workers / paid browser hours | CDP + MCP, WebMCP, Live View, Human-in-the-Loop, session recordings; /json Quick Action extracts schema-shaped data with Workers AI or BYOK (2026-07-01) |
✅ | ❌ | ✅ | 🔗 |
| Browser-use ⭐ | Python / Cloud API | Free OSS / Cloud: 5 free prompts, pay-as-you-go $100 credits | Browser Use CLI 3.0 powered by Browser Harness; browser-use skill installs agent skills for Claude Code, Codex, Cursor, Gemini, OpenCode, and related skill directories; Cloud v3 premium agent supports scheduling, persistent memory, skills, and integrations |
✅ | ✅ | ✅ | 🔗 |
| Browser Harness ⭐ | Python / CDP | Free (OSS) | Thin self-healing CDP harness for direct LLM control of real Chrome; packaged skill for coding agents; pinned by Browser Use CLI 3.0 | ✅ | ❌ | ✅ | 🔗 |
| Vercel agent-browser ⭐ | Rust / Node CLI | Free (OSS) | Browser automation CLI for AI agents; compact text snapshots, ref-based actions, Chrome and Lightpanda engines, packaged agent skill | ✅ | ❌ | ✅ | 🔗 |
| Stagehand v3.6 ⭐ | TypeScript/Python | Free (OSS) | v3.6 (2026-06-19): WebMCP support; Claude Fable 5 support; Microsoft Entra ID auth; full CDP rewrite, 44% faster, action caching, hybrid deterministic + AI | ✅ | ❌ | ✅ | 🔗 |
| LaVague | Python | Free (OSS) | NL to code | ❌ | ❌ | ✅ | 🔗 |
| Skyvern | Python | Free tier / $29–$149/mo | CV-based automation, Ollama support | ✅ | ✅ | ✅ | 🔗 |
| Notte | Python/Cloud | Free tier / $29/mo+ | Deterministic replay, demo→script | ✅ | ❌ | ✅ | 🔗 |
| Firecrawl | Python / CLI | Free tier / $49/mo+ | LLM-powered crawling & scraping | ✅ | ❌ | ✅ | 🔗 |
| Playwright MCP | TypeScript | Free (OSS) | Cross-browser automation, VS Code | ✅ | ❌ | ✅ | 🔗 |
| Langflow | Python | Free (OSS) / Cloud $29/mo | Visual multi-agent & RAG workflows | ✅ | ✅ | ✅ | 🔗 |
| LlamaIndex | Python | Free (OSS) / Cloud $29/mo | Document-heavy RAG, retrieval quality | ✅ | ✅ | ✅ | 🔗 |
| Haystack | Python | Free (OSS) / Cloud $49/mo | Regulated deployments, structured pipelines | ✅ | ✅ | ✅ | 🔗 |
| AgentQL | TypeScript/Python | Free (1K req/mo) / $49/mo / $149/mo | Natural language web querying/automation | ✅ | ✅ | ✅ | 🔗 |
| ScrapeGraphAI | Python | Free OSS / Cloud $29/mo | Natural language web scraping | ✅ | ✅ | ✅ | 🔗 |
| WebVoyager | Python | Free (OSS) | Autonomous web browsing research | ❌ | ❌ | ✅ | 🔗 |
| Anchor Browser | TypeScript, Python, REST | Free credits tier / usage-based plans | Managed browser fleets + AI steps for computer-use-style web tasks; see docs.anchorbrowser.io | ✅ | ✅ | ✅ | 🔗 |
| Hyperbrowser | TypeScript, Python, REST | Free credits + pay-as-you-go (see hyperbrowser.ai/pricing) | Hosted agents: Browser-Use, Claude/OpenAI/Gemini computer use, Stagehand, HyperAgent; optional BYOK for LLM | ✅ | ✅ | ✅ | 🔗 |
| Steel | Python, Node.js, REST | Hobby credits + paid tiers; OSS self-host via Docker (steel.dev, docs.steel.dev) | Sessions API for Playwright/Puppeteer/Selenium; self-hosted steel-browser avoids vendor browser caps; agent cookbooks | ✅ | ✅ | 🔗 |
| Service | Platform | Pricing | Best For | GitHub |
|---|---|---|---|---|
| ChatGPT agent | ChatGPT | Plus $20/mo / Pro $200/mo / Team | Guided browser tasks, research, forms, spreadsheets, code execution | ❌ |
| Gemini Spark ⭐ | Google AI Ultra | Included with Google AI Ultra ($99.99/mo, US beta from 2026-05-26) | 24/7 always-on personal AI agent; runs on Google Cloud VMs while device is off; MCP connections to Canva, OpenTable, Instacart at launch; Adobe, GitHub, Notion, Slack confirmed summer 2026 | ❌ |
| Claude Cowork | Anthropic (desktop app) | Included with Claude paid plans (GA 2026-04-09) | Desktop agent: works in local files & apps, multi-step task completion, Dispatch (assign by text/voice), runs while user is away | ❌ |
| Project Mariner | Google AI Ultra | Included with Google AI Ultra ($99.99/mo) | Multi-step browser tasks, shopping, and reservations | ❌ |
| Skyvern Cloud | Cloud API | Free 1K credits / Hobby $29/mo / Pro $149/mo | Resilient CV-based automation, Ollama support | 🔗 |
| Browserbase | Cloud API | Free credits / paid tiers | Stealth mode, session recording, CDP access | ❌ |
| Amazon Nova Act | API (AWS) | Pay-per-use | Autonomous web agent, multi-step browser workflows | ❌ |
| Amazon Bedrock AgentCore ⭐ | AWS | Pay-per-use | Platform for building, connecting, securing, tracing, and operating production agents with any model or framework; includes Runtime, Gateway, Memory, Browser tool, Code Interpreter, Identity, and Observability | 🔗 |
| Vercel Sandbox ⭐ | Vercel | Usage-based Vercel platform | Agent/runtime sandbox for code execution; FUSE mounts for S3, network, and custom filesystems added 2026-07-03; pairs with AI SDK and eve workflows | 🔗 |
| OpenAI Workspace Agents | ChatGPT Business, Enterprise, Edu, Teachers | Credit-based after preview | Research-preview workplace agents usable in ChatGPT and Slack channels; role-based admin controls | 🔗 |
| Claude Science ⭐ | Anthropic | Applications through 2026-07-15 00:00 UTC | AI workbench for scientists; auditable artifacts, package/tool integration, flexible compute access, up to $30K Claude credits for selected projects | 🔗 |
| Databricks Genie One ⭐ | Databricks | Free $10/user/mo | GA 2026-06-16; data-smart agentic coworker; iOS/Android/web; 50+ app integrations; Genie Ontology context layer | ❌ |
| Thoughtworks Agent/works ⭐ | Multi-cloud | Enterprise | GA 2026-06-16; governed runtime for enterprise agents; single control plane; any cloud deployment | ❌ |
| Kore.ai Artemis ⭐ | Azure (initial) | Enterprise | GA 2026-05-21; AI-native agent platform; governance + observability; 40+ channels; Microsoft Foundry integration | ❌ |
| Huawei Cloud AgentArts ⭐ | Huawei Cloud | Enterprise | GA 2026-06-08; enterprise-grade agent platform; open-source edition (openJiuwen); AgentArts Orchard portal | ❌ |
| Zensar ZenseAI.AgentMesh ⭐ | Multi-cloud | Enterprise | GA 2026-06-19; 80+ pre-built agents; governed orchestration; EU AI Act alignment | ❌ |
| Sedai AI Agent Optimization ⭐ | Multi-cloud | Early access | GA planned late 2026; smart model routing; governance + observability; supports OpenAI/Bedrock/Vertex/Azure | ❌ |
| GMI Cloud AgentBox ⭐ | GMI Cloud | Usage-based | GA 2026-06-08; marketplace for AI agents; 100+ models via single key; MaaS inference | ❌ |
| ClickHouse Agents ⭐ | ClickHouse | Free (public beta) | GA 2026-06-09; Claude-powered agentic analytics; no-code agent builder; MCP-compatible | ❌ |
| Google Data Agents ⭐ | Google Cloud | Free (preview) | GA 2026-06-16; Data Science Agent, Database Agent, Looker Dashboard Agent, Data Insights Agent, Deep Research Agent | ❌ |
| C1 Autonomous Worker ⭐ | C1 | Enterprise | GA 2026-06-15; agent that runs work loops, carries state, executes code; available via Slack | ❌ |
Control a computer or cloud sandbox using plain English text — no coding required. Just describe the task and the agent handles everything: clicking, typing, navigating, running code, and completing multi-step workflows.
Legend: 🖥️ = runs on your physical computer | ☁️ = cloud/sandbox computer | 🌐 = controls a browser | 🔁 = multi-agent/parallel | 💬 = simple English prompt | 🔓 = free/open-source | 💰 = paid
These services run in a cloud sandbox (virtual Linux/Windows desktop), control the computer for you, and are driven purely by natural language instructions.
| Agent | Interface | Pricing | Multi-Agent | Parallel Sessions | Local LLM | English Prompt | GitHub |
|---|---|---|---|---|---|---|---|
| Suna (Kortix) | Web dashboard | Free (OSS self-host); Cloud from $20/mo; Pro $29/mo; Team $199/mo | ✅ | ✅ | ❌ | ✅ | 🔗 |
| OpenManus | Web dashboard | Free OSS | ✅ | ✅ | ❌ | ✅ | 🔗 |
| Manus AI | Web dashboard | Free (300 credits/day) / Plus $20/mo / Pro $200/mo | ✅ | ✅ | ❌ | ✅ | ❌ |
| ChatGPT Agent | ChatGPT Web/App | Plus $20/mo / Pro $200/mo | ❌ | ❌ | ❌ | ✅ | ❌ |
| Gemini Computer Use | API / AI Studio | Gemini Pro $19.99/mo / API metered | ❌ | ❌ | ❌ | ✅ | ❌ |
| Devin | Web dashboard | Core $20/mo ($2.25/ACU) / Team $500/seat/mo | ❌ | ❌ | ❌ | ✅ | ❌ |
| OpenHands | Web UI / CLI | Free OSS / Cloud Individual free | ✅ | ✅ | ✅ (any API) | ✅ | 🔗 |
| E2B Desktop Sandbox | API / SDK | Hobby free / Pro $150/mo | ❌ | ✅ (via code) | ✅ | ✅ | 🔗 |
| Cua (trycua) | CLI / Python SDK | Free (OSS) | ❌ | ✅ | ✅ | ✅ | 🔗 |
| Airtop | Web dashboard / API | Starter $26/mo (3 sessions) / Pro $80/mo (30 sessions) | ❌ | ✅ | ❌ | ✅ | ❌ |
| Skyvern Cloud | Web dashboard / API | Free 1K credits / Hobby $29/mo / Pro $149/mo | ❌ | ✅ | ✅ (Ollama) | ✅ | 🔗 |
| Convergence Proxy | Web / API | Free tier / Pro $20/mo (acquired by Salesforce) | ❌ | ❌ | ❌ | ✅ | ❌ |
| Amazon Nova Act | API (AWS) | Pay-per-use (AWS pricing) | ❌ | ✅ | ❌ | ✅ | ❌ |
| Project Mariner | Google AI Ultra | Included ($100/mo Ultra 5× plan) | ❌ | ❌ | ❌ | ✅ | ❌ |
| Perplexity Computer | Web dashboard | Perplexity Pro $20/mo | ❌ | ❌ | ❌ | ✅ | ❌ |
| OpenAI Computer Use (API) | API / ChatGPT | $15/M input, $60/M output | ✅ | ✅ | ❌ | ✅ | ❌ |
| OpenHands Agent Canvas ⭐ | Web / Self-hosted | Free (OSS) | Workspace for creating automations integrating Slack/GitHub; self-hostable; released 2026-06-16 | 🔗 |
These agents run on your own machine, see your screen, and control your keyboard/mouse — no cloud required.
| Agent | Windows | macOS | Linux | Dashboard/UI | CLI | API/LLM | Multi-Agent | Parallel Sessions | Pricing | GitHub |
|---|---|---|---|---|---|---|---|---|---|---|
| Claude Computer Use | ✅ | ✅ | ✅ | ❌ | ✅ (API) | Claude API ($3–$15/M) | ❌ | ❌ | Claude API ($3–$15/M tokens) | Commercial |
| Agent TARS (ByteDance) | ✅ | ✅ | ✅ | ✅ Web UI | ✅ npx @agent-tars/cli@latest |
Any LLM | ❌ | ✅ | Free (OSS) | 🔗 |
| Aiden | ❌ | ❌ | ✅ | ✅ Desktop | ✅ | Local / BYO API | ❌ | ❌ | Free (OSS) | 🔗 |
| Simular Agent S2 | ✅ | ✅ | ✅ | ❌ | ✅ | Any LLM API | ❌ | ❌ | Free (OSS) | 🔗 |
| UI-TARS Desktop (ByteDance) | ✅ | ✅ | ✅ | ✅ Desktop app | ❌ | UI-TARS-2 model | ❌ | ❌ | Free (OSS) | 🔗 |
| Open Interpreter | ✅ | ✅ | ✅ | ✅ Web | ✅ interpreter |
Any (OpenAI, Claude, local) | ❌ | ❌ | Free (OSS) | 🔗 |
| Open-Interface | ✅ | ✅ | ✅ | ❌ | ✅ | GPT-4V / any vision LLM | ❌ | ❌ | Free (OSS) | 🔗 |
| Agent S / S2 | ✅ | ✅ | ✅ | ❌ | ✅ | Any LLM API | ❌ | ❌ | Free (OSS) | 🔗 |
| UFO (Microsoft) | ✅ | ❌ | ❌ | ✅ UI | ✅ | GPT-4V / Azure | ❌ | ❌ | Free (OSS) | 🔗 |
| Windows-Use | ✅ | ❌ | ❌ | ❌ | ✅ | Any vision LLM | ❌ | ❌ | Free (OSS) | 🔗 |
| Bytebot | ❌ | ❌ | ✅ | ✅ (Docker) | ✅ | Any LLM | ❌ | ❌ | Free (OSS) | 🔗 |
| OpenCUA | ✅ | ✅ | ✅ | ❌ | ✅ | Any | ❌ | ❌ | Free (OSS) | 🔗 |
| Windows Agent Framework (Microsoft) | ✅ | ❌ | ❌ | Agent Designer in VS 2026 | ✅ wagent |
Any LLM | ❌ | ❌ | Free (MIT) | 🔗 |
| Khoj | ✅ | ✅ | ✅ | ✅ Web UI | ✅ | Any (Ollama, LM Studio, OpenAI) | ❌ | ❌ | Free (OSS) / Cloud $10/mo | 🔗 |
| Eigent (CAMEL-AI) | ✅ | ✅ | ✅ | ❌ | ✅ | Any LLM | ✅ | ✅ | Free (OSS) | 🔗 |
Control a browser with natural language — click, fill forms, scrape, automate. No script writing needed.
| Agent | Type | Pricing | Dashboard | CLI | Multi-Agent | Parallel Sessions | Local LLM | GitHub |
|---|---|---|---|---|---|---|---|---|
| Browser-use ⭐ | OSS Python lib + Cloud | Free OSS / Cloud: 5 free prompts, pay-as-you-go $100 credits; v3 premium agent | ✅ Cloud | ✅ CLI 3.0 | ✅ | ✅ | ✅ (Ollama) | 🔗 |
| Stagehand | OSS TypeScript | Free (OSS) | ❌ | ✅ | ❌ | ✅ | ✅ | 🔗 |
| NanoBrowser | Chrome extension | Free (OSS) | ✅ Extension | ❌ | ✅ | ❌ | ✅ (Ollama) | 🔗 |
| Skyvern | Python / Cloud | Free tier / $29–$149/mo | ✅ Cloud | ✅ | ✅ | ✅ | ✅ (Ollama) | 🔗 |
| Openator | Python | Free (OSS) | ❌ | ✅ | ❌ | ✅ | ✅ | 🔗 |
| Open Operator | Web UI | Free | ✅ | ❌ | ❌ | ❌ | ❌ | 🔗 |
| Airtop | Web / API | $26–$80/mo | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ |
| MultiOn | API / Chrome ext | Free / Paid | ✅ | ❌ | ✅ | ❌ | ❌ | 🔗 |
Coordinate multiple AI agents in parallel to complete complex workflows — driven by plain English goals.
| Platform | Type | Dashboard | CLI | Cloud | Local LLM | Parallel | Pricing | GitHub |
|---|---|---|---|---|---|---|---|---|
| CrewAI | Multi-agent OSS + Cloud | ✅ AMP Studio | ✅ crewai |
✅ AMP | ✅ | ✅ | Free OSS / Starter $99/mo / Pro $299/mo / Enterprise custom | 🔗 |
| Microsoft Agent Framework (AutoGen + Semantic Kernel unified) | Multi-agent conversations (v1.0 GA April 2026) | ✅ AG2 Studio | ✅ Python | ✅ Azure | ✅ | ✅ | Free (OSS) / Azure pay-per-token | 🔗 |
| LangGraph | Stateful agent graphs | ✅ LangSmith | ✅ Python | ✅ Cloud | ✅ | ✅ | Free OSS / Professional $99/mo | 🔗 |
| OpenHands | Dev-focused multi-agent | ✅ Web UI | ✅ | ✅ Cloud | ✅ | ✅ | Free (OSS + Cloud free tier) | 🔗 |
| OWL (Camel-AI) | Distributed multi-agent | ❌ | ✅ Python | ❌ | ✅ | ✅ | Free (OSS) | 🔗 |
| Manus AI | Cloud multi-agent | ✅ Web | ❌ | ✅ | ❌ | ✅ | Free 300 credits/day / $20–$200/mo | ❌ |
| n8n | Workflow + AI agents | ✅ Visual canvas | ✅ n8n |
✅ Cloud | ✅ (Ollama node) | ✅ | Free OSS / Starter $24/mo / Pro $60/mo | 🔗 |
| Devin | Software engineering | ✅ Web | ❌ | ✅ | ❌ | ❌ | Core $20/mo ($2.25/ACU) / Team $500/seat | ❌ |
| Smolagents (HuggingFace) | Lightweight code agents | ❌ | ✅ Python | ❌ | ✅ | Free (OSS) | 🔗 | |
| C3 Code (C3 AI) | Enterprise AI dev platform | ✅ Web | ❌ | ✅ | ❌ | ✅ | Enterprise pricing | Full SDLC agents, governed deployment, natural language to production apps |
| Dify | Visual LLM platform | ✅ Web UI | ✅ | ✅ Cloud | ✅ | ✅ | Free OSS / Cloud plans | 🔗 |
| OpenAI Agents SDK | Agent handoffs + tool orchestration | ❌ | ✅ Python / TS | ✅ OpenAI API | ❌ | ✅ | Free (OSS) / OpenAI API costs | 🔗 |
| Google ADK | Hierarchical agent tree, A2A protocol | ✅ Vertex AI | ✅ Python | ✅ Vertex AI | ❌ | ✅ | Free (OSS) / Vertex AI pay-per-token | 🔗 |
| ADK Go 2.0 ⭐ | Graph-based agent workflow engine | — | ✅ go run |
✅ Google Cloud | ❌ | ✅ | Free (OSS) | 🔗 |
| Genkit Agents ⭐ | Full-stack agent framework | ✅ Developer UI | — | — | ✅ | ✅ | Free (OSS) | 🔗 |
| AI SDK 7 ⭐ | Production agent platform (TypeScript) | — | ✅ @ai-sdk/tui |
✅ Vercel | ✅ | ✅ | Free (OSS) | 🔗 |
| Mastra ⭐ | Agent framework with file-based agents, skills, workspace, and subagents | — | ✅ mastra |
✅ Vercel | ✅ | ✅ | Free (OSS) | 🔗 |
| Ruflo | Multi-agent orchestration | ❌ | ✅ Python | ❌ | ✅ | ✅ | Free OSS / Cloud plans | 🔗 |
| Vercel eve ⭐ | Open-source agent framework with Agent Runs traces via Vercel MCP/CLI | ✅ Web | ✅ vercel agent-runs |
✅ Vercel | ✅ | ✅ | Free (OSS) / Vercel hosting costs | 🔗 |
| Omnigent (Databricks) ⭐ | Meta-harness for agents | ✅ Web/App | ✅ | ✅ | ✅ | ✅ | Free (Apache 2.0) | 🔗 |
| GitKraken Kepler ⭐ | Agentic Development Environment | ✅ Desktop | ✅ | ❌ | ✅ | ✅ | Free tier / paid plans | 🔗 |
| LangChain Deep Agents ⭐ | Long-horizon agent harness | ❌ | ✅ Python | ❌ | ✅ | ✅ | Free (MIT) | 🔗 |
| Wheelie (Continua) ⭐ | Agentic dev OS + service runtime | ✅ Web | ✅ | ✅ VMs | ✅ | ✅ | Free (waiting list) | 🔗 |
| Cloudflare Flue ⭐ | Agent framework on Workers | ✅ Dashboard | ✅ | ✅ Cloudflare | ❌ | ✅ | Free (OSS) / Workers pricing | 🔗 |
| C1 Autonomous Worker ⭐ | Enterprise agent | ✅ (Slack) | ❌ | ✅ | ❌ | ❌ | Enterprise | ❌ |
| ClickHouse Agents ⭐ | Agentic analytics | ✅ | ❌ | ✅ | ❌ | ✅ | Free (public beta) | ❌ |
| Google Data Agents ⭐ | Data agents | ✅ | ❌ | ✅ | ❌ | ✅ | Free (preview) | ❌ |
| Salesforce Agentforce 3 ⭐ | MCP-native agent platform | ✅ Web | ❌ | ✅ | ❌ | ✅ | Enterprise pricing | ❌ |
| Kore.ai Artemis ⭐ | AI-native agent platform | ✅ | ❌ | ✅ (Azure) | ❌ | ✅ | Enterprise | ❌ |
| Konecta Kolibri ⭐ | Agentic AI orchestration | ✅ | ❌ | ✅ | ❌ | ✅ | Enterprise | ❌ |
| Zensar ZenseAI.AgentMesh ⭐ | Enterprise agentic AI | ✅ | ❌ | ✅ | ❌ | ✅ | Enterprise | ❌ |
| Sedai AI Agent Optimization ⭐ | Agent optimization | ✅ | ❌ | ✅ | ❌ | ✅ | Early access | ❌ |
Tools supporting parallel agent orchestration (✅) vs single-agent only (❌):
| Category | Supports Parallel Agents | Tools |
|---|---|---|
| Cloud Sandbox | ✅ | Manus AI, OpenHands, E2B Desktop Sandbox, Cua (trycua), Airtop, Skyvern Cloud, Amazon Nova Act, Perplexity Computer, OpenAI Computer Use (API) |
| Cloud Sandbox | ❌ | ChatGPT Agent, Gemini Computer Use, Devin, Convergence Proxy, Project Mariner |
| Local Machine | ✅ | Agent TARS, E2B Desktop Sandbox, Cua (trycua) |
| Local Machine | ❌ | Claude Computer Use, UI-TARS Desktop, Open Interpreter, Open-Interface, Agent S/S2, UFO, Windows-Use, Bytebot, OpenCUA, Khoj |
| Browser-Only | ✅ | Browser-use, Skyvern, Airtop, MultiOn, Browser Operator |
| Browser-Only | ❌ | Stagehand, NanoBrowser, Openator, Open Operator |
| Developer Libraries | ✅ | Browser-use, Browser Harness, Vercel agent-browser, Skyvern, Cloudflare Browser Run, Langflow, LlamaIndex, Haystack, AgentQL, ScrapeGraphAI, WebVoyager, Hyperbrowser, Anchor Browser, Steel |
| Developer Libraries | ❌ | Chrome DevTools MCP, Stagehand, LaVague, Notte, Firecrawl, Playwright MCP |
| Multi-Agent Platforms | ✅ | CrewAI, Microsoft Agent Framework, LangGraph, OpenHands, OWL, Manus AI, n8n, Smolagents, Dify, OpenAI Agents SDK, Google ADK, ADK Go 2.0, Genkit Agents, AI SDK 7, Mastra, Vercel eve |
| Multi-Agent Platforms | ❌ | Devin |
Tools, frameworks, and specialized models for building production AI systems — from embeddings and video generation to safety, evaluation, and model routing.
Specialized models for converting text (or images) into dense vector representations and for reranking retrieval results. Essential infrastructure for RAG pipelines and semantic search. Prices as of June 2026.
| Model | Developer | Dimensions | Max Tokens | Pricing | Best For | GitHub |
|---|---|---|---|---|---|---|
| Gemini Embedding 2 | 3,072 | 8,192 text / 6 images / 120s video / 180s audio | $0.20/1M text tokens; GA 2026-04-22 | First natively multimodal embedding — text, image, video, audio, PDF in one space | — | |
| Granite Embedding Multilingual R2 (311M) ⭐ | IBM | 768 | 32,768 | Free (Apache 2.0) | MTEB Multilingual Retrieval 65.2 (#2 open <500M); 200+ languages; 32K context (64x R1) | 🔗 |
| Granite Embedding Multilingual R2 (97M) ⭐ | IBM | 384 | 32,768 | Free (Apache 2.0) | Best sub-100M multilingual embedder (MTEB 60.3); 200+ languages | 🔗 |
| pplx-embed-v1 (4B) | Perplexity | 3,072 | 8,192 | API pay-per-use | MTEB Multilingual leader; web-scale retrieval; contextual variant available | 🔗 |
| LFM2.5-Embedding-350M ⭐ | Liquid AI | 1,024 | 8,192 | Free (open-source) | 350M-param bidirectional embedder; 11 languages; best-in-class multilingual; runs on CPU/edge | 🔗 |
| LFM2.5-ColBERT-350M ⭐ | Liquid AI | Per-token vectors | 8,192 | Free (open-source) | First bidirectional ColBERT; word-by-word matching for higher accuracy; 11 languages; runs on CPU/edge | 🔗 |
| ML-Embed-0.6B (CodeFuse) ⭐ | CodeFuse AI | 1,024 | 8,192 | Free (Apache 2.0) | 3D Matryoshka Learning; multilingual; ICML 2026 paper | 🔗 |
| Qwen3-Embedding-8B | Alibaba / Qwen | 4,096 (flex) | 32K | Free (Apache 2.0) | MTEB leader (70.6); 100+ languages, instruction-aware, flexible dimensions; 0.6B/4B/8B sizes | 🔗 |
| jina-embeddings-v5-omni-small ⭐ | Jina AI / Elastic | 1,024 (Matryoshka 32-1,024) | 32K text; image, video, audio, PDF | API pay-per-use | Multimodal embedding model for text, image, video, audio, and PDF search; frozen v5 text backbone keeps text outputs aligned with v5-text-small | 🔗 |
| jina-embeddings-v4 | Jina AI | 32–2,048 (flex) | 8,192 | API pay-per-use | Built on Qwen2.5-VL-3B (3.8B); 3 LoRA adapters (query/passage/matching) | 🔗 |
| text-embedding-3-small | OpenAI | 1,536 | 8,191 | $0.02/1M tokens | Cost-effective English embeddings | — |
| text-embedding-3-large | OpenAI | 3,072 | 8,191 | $0.13/1M tokens | Highest-quality English retrieval | — |
| Embed v4 | Cohere | 1,536 | 128K | $0.12/1M (text), $0.47/1M (image) | Multimodal text + image RAG | — |
| voyage-3-large | Voyage AI | 256–2,048 (flex) | 32K | ~$0.18/1M tokens | Highest-quality retrieval, long context | — |
| BGE-M3 | BAAI | 1,024 | 8,192 | Free (open-source) | Multi-functional: dense + sparse + ColBERT | 🔗 |
| Nomic Embed v2 (MoE) | Nomic AI | 256–768 (flex) | 512 | Free (open-source) | Multilingual, MoE efficiency (305M active) | 🔗 |
| text-embedding-005 | Google (Vertex AI) | 768 | 2,048 | $0.10/1M tokens | GCP-native semantic search | — |
| Model | Developer | Max Tokens | Pricing | Best For | GitHub |
|---|---|---|---|---|---|
| Rerank 4.0 Pro | Cohere | 32K | $1.00/1K queries | High-accuracy domain-specific reranking | — |
| Rerank 4.0 Fast | Cohere | 32K | $0.50/1K queries | Low-latency production reranking | — |
| rerank-2.5 | Voyage AI | 32K | API pay-per-use | Instruction-following, multilingual | — |
| Qwen3-Reranker-4B | Alibaba / Qwen | 32K | Free (Apache 2.0) | MTEB-R 69.76, MMTEB-R 72.74; 100+ languages; instruction-following | 🔗 |
| BGE Reranker v2-m3 | BAAI | 8,192 | Free (open-source) | Open-source cross-encoder reranking | 🔗 |
| Jina Reranker v2 | Jina AI | 8,192 | API pay-per-use | Multilingual, long-context reranking | — |
Text-to-video and image-to-video generation models for creating short clips from prompts. The field is moving rapidly — resolutions, durations, and pricing change frequently. Specs as of June 2026.
| Model | Developer | Resolution | Duration | Pricing | Open Source | Best For | GitHub |
|---|---|---|---|---|---|---|---|
| OpenAI | Up to 1080p | Up to 20s | App shut down 2026-04-26; API shutdown 2026-09-24 | No | Deprecated — use Veo 3 or Luma Ray3 | — | |
| Veo 3.1 | Google DeepMind | 1080p | Up to 8s (extendable) | ~$0.20–$0.40/s | No | Native 48kHz audio + video, realistic physics; only model with synced dialogue | — |
| Gemini Omni Flash ⭐ | Google DeepMind | 1080p | Public preview | $0.10/s | No | High-quality video generation + conversational editing; text/image/video input; released 2026-06-30 | — |
| Seedance 2.0 | ByteDance | 1080p | Up to 16s | API pay-per-use | No | #1 Artificial Analysis leaderboard (Elo 1213, June 2026); accepts 9 images + 3 clips + 3 audio | — |
| Luma Ray3.14 | Luma AI | Up to 1080p | Up to 10s | $29–$99/mo | No | Native 1080p, 4× faster at 720p vs Ray2, 3× lower cost | — |
| LTX-2 | Lightricks | Up to 4K | Up to 20s @ 50fps | Free (open-source; commercial license for >$10M ARR) | Yes (Apache 2.0) | First open-source 4K audio+video sync model, 19B DiT params, runs on consumer GPU | 🔗 |
| Runway Gen-4 / Gen-4.5 | Runway | Up to 4K | Up to 16s | $12–$76/mo | No | Professional creative workflows | — |
| Kling 3.0 | Kuaishou | 1080p | Up to 15s | Free / $5.99–$66/mo | No | Multi-Shot Storyboard, best value-for-money; released 2026-02-04 | — |
| Pika 2.0 | Pika Labs | 1080p | Up to 5s | Free / $8–$58/mo | No | Social media, creative effects | — |
| MiniMax Video-01 | MiniMax | 720p | Up to 6s | ~$0.40/video | No | Strong text-motion responsiveness | — |
| HunyuanVideo | Tencent | 720p–2K | Up to 16s | Free (self-host; ~60GB VRAM) | Yes (Apache 2.0) | High per-frame fidelity, long clips | 🔗 |
| Wan 2.2 (14B) | Alibaba | 480p–1080p | Up to 10s | ~$0.10–$0.30/clip (API) | Yes (Apache 2.0) | Motion quality, VBench #1 benchmark | 🔗 |
| MiniMax Hailuo 2.3 ⭐ | MiniMax | 1080p | Up to 10s | API pay-per-use | No | Released 2026-06-22; successor to Hailuo 02 | — |
| MiniMax Hailuo 02 ⭐ | MiniMax | 1080p | Up to 10s | API pay-per-use | No | 3x params vs predecessor, 4x training data, best complex instruction adherence; #2 Artificial Analysis Video Arena; released 2026-06-20 | — |
| Kling 3.0 Turbo ⭐ | Kuaishou | 480p–720p | 1–15s preview | Free / paid | No | Fast-preview mode for rapid iteration; released 2026-06-17 | — |
| MiniMax Video-01 ⭐ | MiniMax | 720p | Up to 6s | ~$0.40/video | No | First AI-native video model from MiniMax; text-to-video + image-to-video; released 2026-06-20 | — |
| Grok Imagine Video 1.5 ⭐ | xAI | 720p | Up to 15s | $0.06/s ($4.20/min at 720p) | No | #1 Image-to-Video Arena leaderboard (Elo 1,473); native audio+video sync; Fast variant 25s for 6s clip; GA 2026-06-16 | — |
| HappyHorse-1.0 ⭐ | Alibaba (ATH) | 1080p | Up to 5s | Free (open-source) | Yes (open-source) | #1 on Artificial Analysis video leaderboard; unified Transformer; simultaneous audio/video; 15B params; API coming 2026-04-30 | 🔗 |
| Seedance 2.1 ⭐ | ByteDance | 1080p | Up to 16s | API pay-per-use | No | Sharper motion, tighter character consistency; available on Atlas Cloud Day 0 | — |
| Microsoft Mirage ⭐ | Microsoft Research | — | — | Free (research) | Research | Video world model with persistent spatial memory; 10.57x faster generation; built on Wan2.2 | 🔗 |
| MoVerse ⭐ | Academic/MSFT | — | — | Free (research) | Research | Real-time video world model; panoramic Gaussian scaffold; 8 FPS on RTX 4090 | — |
| Echo-Infinity ⭐ | JD/Echo Team | — | 24h+ rollouts | Free (research) | Research | Real-time infinite video generation; learnable evolving memory; 1.3M+ frames | 🔗 |
| JoyAI-Echo ⭐ | JD | Up to 1080p | Minute-level | Free (research) | Research | Multi-shot audio-video generation; 7.5x speedup via DMD distillation | 🔗 |
| Mochi 1 | Genmo | 480p | Up to 5.4s @ 30fps | Free (open-source) | Yes (Apache 2.0) | High-quality open text-to-video | 🔗 |
| CogVideoX | Zhipu AI / Tsinghua | 720p | ~6s | Free (open-source) | Yes (Apache 2.0) | Image-to-video quality, LoRA fine-tuning | 🔗 |
Text-to-speech (TTS) and speech-to-text (STT / ASR) models for voice generation, transcription, and real-time audio. Prices as of July 2026.
| Rank | Model | Elo | Pricing | Languages |
|---|---|---|---|---|
| 🥇 #1 | Sonic-3.5 ⭐ | ~1,220 | ~$39 / 1M chars | 40+ |
| 🥈 #2 | Gemini 3.1 Flash TTS | ~1,216 | $18.30 / 1M chars | 70+ |
| 🥉 #3 | Realtime TTS-2 (Research Preview) | ~1,208 | $25–$35 / 1M chars | 100+ |
| #4 | Sonic 4 | ~1,210 | ~$46.70 / 1M chars | 40+ |
| #5 | Realtime TTS 1.5 Max | ~1,200 | $35 / 1M chars | 100+ |
Arena Elo scores shift continuously. Treat rankings as point-in-time readings. Source: Artificial Analysis, June 2026. Cartesia positions Sonic-3.5 and Ink-2 as its latest production speech stack for real-time voice agents. Source: 🔗
| Model | Developer | Languages | Real-time | Open Source | Pricing | Best For | GitHub |
|---|---|---|---|---|---|---|---|
| Sonic-3.5 ⭐ | Cartesia | 42 | Yes (sub-90ms) | No | See Cartesia pricing | Fastest, most natural Cartesia TTS model; ranked #1 for naturalness; GA in 2026 changelog | 🔗 |
| Gemini 3.1 Flash TTS | 70+ | Yes | No | $18.30 / 1M chars | Audio tags for granular style/pace control, SynthID watermarking | — | |
| Realtime TTS-2 (Research Preview) | Inworld AI | 100+ | Yes | No | $25–$35 / 1M chars | Realtime conversation, cross-lingual voice identity, emotional perception | — |
| Sonic 4 ⭐ | Cartesia | 40+ | Yes | No | ~$46.70 / 1M chars | Sonic 4 Turbo ~40ms TTFA (May 2026); 40+ languages, 95% world pop; instant 3s voice clone | — |
| Realtime TTS 1.5 Max | Inworld AI | 100+ | Yes | No | $35 / 1M chars | Realtime conversational agents, low latency + low cost | — |
| Eleven v3 | ElevenLabs | 70+ | No | No | Subscription / API (up to 55% price cut May 2026) | Most expressive TTS, inline audio tags [whispers], [laughs], multi-speaker dialogue |
— |
| Eleven Flash v2.5 | ElevenLabs | 32 | Yes (~75ms) | No | Subscription / API | Ultra-fast real-time, same voice library as offline | — |
| Eleven Multilingual v2 | ElevenLabs | 30 | Yes | No | Subscription / API | Emotionally-aware multilingual synthesis | — |
| gpt-4o-mini-tts | OpenAI | 50+ | Yes | No | $0.60 / $12 per 1M (in/out) | Natural-language voice steering, 13 built-in voices | — |
| GPT-Realtime-2 | OpenAI | — | Yes | No | — | Speech-to-speech with GPT-5-class reasoning, tool calls, interruptions — released 2026-05-07 | — |
| GPT-Realtime 2.1 / 2.1-mini ⭐ | OpenAI | — | Yes | No | — | Newer Realtime API speech-to-speech models with function calling, MCP servers, and SIP; shipped 2026-07-06 | 🔗 |
| gpt-realtime-mini | OpenAI | — | Yes | No | — | Cost-efficient speech-to-speech | — |
| GPT-Live-1 / 1 mini ⭐ | OpenAI | — | Yes (full-duplex) | No | Bundled in ChatGPT tiers (API pending) | Consumer voice model powering ChatGPT Voice; full-duplex, listens while speaking; GPT-5.5 backend delegates complex work; announced 2026-07-08, API "soon" | 🔗 |
| gpt-audio-mini | OpenAI | — | Yes | No | — | Cost-efficient audio generation | — |
| OpenAI TTS / TTS HD | OpenAI | 57 | Yes | No | $15 / $30 per 1M chars | Enterprise, seamless GPT integration | — |
| Grok TTS | xAI | 25+ | Yes | No | $15.00 / 1M chars | Speech tags, 80+ voices, SOC 2 / HIPAA compliant | — |
| Deepgram Aura-2 | Deepgram | 7 | Yes (~90–200ms) | No | $0.030 / 1K chars ($30/1M) | Enterprise voice agents, unified STT+TTS stack, on-prem deployment | — |
| Hume Octave 2 | Hume AI | 11+ | Yes (~100–200ms) | No | Varies (contact sales) | Emotional intelligence, voice conversion, phoneme editing | — |
| Supertonic 3 ⭐ | Supertone | 31 | Yes | No | Subscription / API | Fast, cost-efficient multilingual narration | — |
| Lightning V3.1 / V3.2 | Smallest.ai | 15 | Yes | No | Pay-as-you-go | Conversational TTS, MOS 3.89, auto language detection, mid-sentence switching | — |
| StepAudio 2.5 Realtime ⭐ | StepFun | Chinese, English | Yes | No | — | End-to-end real-time speech LLM, paralinguistic comprehension, persona RLHF | — |
| VibeVoice-Realtime-0.5B ⭐ | Microsoft | English (9 experimental langs) | Yes (~300ms) | Yes (MIT) | Free (self-host) | 0.5B real-time TTS; streaming text input; 10-min robust generation; Qwen2.5 base; <300ms TTFA; released 2025-12-03; HuggingFace integration 2026-03-06 | 🔗 |
| Qwen3-TTS 🇨🇳 | Alibaba | 10 | Yes (streaming) | Yes (Apache 2.0) | Free (self-host) / API | Voice design, voice cloning, instruction control, multilingual | 🔗 |
| Stability Audio 3.0 | Stability AI | Music/SFX | Yes (small/medium OSS) | Yes (small/medium, Apache 2.0) | Free (small/medium OSS weights); commercial via API | Professional-grade music >6 min, open-weight small/medium variants | 🔗 |
| Sesame CSM | Sesame AI Labs | English | Yes | Yes | Free | Conversational, emotionally expressive (4.7 MOS) | 🔗 |
| MAI-Voice-2 ⭐ | Microsoft | 15 | Yes | No | API (Foundry) | Most expressive Microsoft TTS; granular emotion control (whisper/sad/etc.); 15 languages; custom voice from 5-60s clip; released 2026-06-02 | — |
| Higgs Audio v3 TTS ⭐ | Boson AI | 100+ | Yes | No | API / free (self-host) | Voice chat optimized; inline emotion/style/prosody tags; zero-shot voice cloning; released 2026-06-04 | 🔗 |
| Chatterbox Multilingual v3 ⭐ | Resemble AI | 25 | Yes | Yes (MIT) | Free (self-host) / API | 0.5B Llama backbone; PerTh watermarking; improved speaker similarity; released 2026-06-10 | 🔗 |
| MisoTTS ⭐ | Miso Labs | — | Yes | Yes (Modified MIT) | Free (self-host) / API pending | 8B emotive TTS; RVQ scales vocabulary; conditions on text + audio; released 2026-06-03 | 🔗 |
| MiniMax Speech 2.8 | MiniMax | 40+ | Yes | No | API pay-per-use | Native sound tags, high-fidelity cloning, studio-grade clarity; introduced 2026-01-23 00:00 UTC | 🔗 |
| ZONOS2 ⭐ | Zyphra | Multilingual | Yes | Yes (Apache 2.0) | Free (self-host) / API | 8B MoE TTS; first open-source MoE TTS; multilingual + code-switched; zero-shot voice cloning; ECAPA-TDNN speaker embeddings; released 2026-06-12 | 🔗 |
| Mistral Voxtral TTS ⭐ | Mistral AI | 9 | Yes | Yes (CC BY-NC 4.0) | $0.016/1K chars | 4B parameter TTS; 70ms latency; voice cloning from 3s audio; released 2026-06-18 | 🔗 |
| MiniMax Speech-01-HD ⭐ | MiniMax | 17 | Yes | No | API | 300+ pre-built voices; high-fidelity voice cloning from 10s audio; released 2026-06-19 | — |
| dots.tts ⭐ | RedNote (Xiaohongshu) | — | Yes | Yes (Apache 2.0) | Free | 2B fully continuous AR TTS; 48kHz AudioVAE; no discrete tokens; released 2026-06 | 🔗 |
| Model | Developer | Languages | Real-time | Open Source | Pricing | Best For | GitHub |
|---|---|---|---|---|---|---|---|
| Ink-2 ⭐ | Cartesia | 40+ | Yes (~100ms) | No | Credit-based | Real-time transcription model paired with Sonic-3.5 for voice-agent workflows; launched 2026-06-30 00:00 UTC | 🔗 |
| MAI-Transcribe-1.5 ⭐ | Microsoft | 43 | Yes | No | API (Foundry) | SOTA on FLEURS; #3 Artificial Analysis; 5x faster than Gemini 3.1; keyword biasing; 1hr audio in <15s; released 2026-06-02 | — |
| Cohere Transcribe ⭐ | Cohere | 14 | No | Yes (Apache 2.0) | Free (self-host) / Model Vault | #1 HuggingFace Open ASR Leaderboard (5.42% WER); 2B Conformer; released 2026-03-26 | |
| VibeVoice-ASR-7B ⭐ | Microsoft | 50+ | Yes (~15s for 60min audio) | Yes (MIT) | Free (self-host) | 60-min single-pass ASR; speaker diarization + timestamps; 50+ languages; structured Who/When/What output; released 2026-01-21; HuggingFace transformers integration 2026-03-06 | 🔗 |
| Cohere Transcribe v2 ⭐ | Cohere | 14 | Yes | Yes (Apache 2.0) | Free (self-host) | Real-time streaming STT; open-source; released 2026-05 | 🔗 |
| Gladia Solaria-3 ⭐ | Gladia | 5 (EU) | Yes | No | API | #1 on business audio; optimized for noisy/real-world European languages; released 2026-06-10 | — |
| Speechmatics Melia ⭐ | Speechmatics | 55+ | Yes (preview) | No | From $0.129/hr (10hr free) | Code-switching across 55+ languages; lowest-priced Speechmatics model; released 2026-06-17 | — |
| Voxtral Realtime ⭐ | Mistral AI | 13 | Yes | Yes (Apache 2.0) | Free (self-host) | Natively streaming ASR; 480ms delay matches Whisper quality; 4B params; released 2026-02 | 🔗 |
| Soniox v5 Async ⭐ | Soniox | — | No | No | API | Structured speech-to-text; speaker separation; language IDs; normalized structured entities; released 2026-06-11 | — |
| NVIDIA Nemotron 3.5 ASR Streaming 0.6B ⭐ | NVIDIA | 40 | Yes | Yes (Apache 2.0) | Free (self-host) / API | 40 language-locales; language-ID prompt conditioning; punctuation/capitalization; released 2026-06-04 | 🔗 |
| Gnani Prisma v2.5 ⭐ | Gnani AI | 9 (Indian) | Yes | No | API | #1 in 8/9 Indian languages on real-world benchmarks; 15% lower WER for rural Hindi; 14M hours training data; released 2026-06-19 | — |
| Grok STT | xAI | 25+ | Yes | No | $0.10/hr (batch) / $0.20/hr (streaming) | Entity recognition (medical/legal/financial), diarization, multichannel, SOC 2 / HIPAA | — |
| Whisper large-v3 | OpenAI | 100+ | No | Yes (MIT) | $0.006/min (API) | Open-source multilingual baseline | 🔗 |
| GPT-4o Transcribe | OpenAI | 50+ | Yes | No | $0.006/min | High-accuracy managed STT | — |
| GPT-4o-mini-transcribe | OpenAI | 50+ | Yes | No | — | Cost-efficient managed STT | — |
| Deepgram Nova-3 | Deepgram | 36+ | Yes | No | $0.0043/min | Ultra-low latency, production STT | — |
| Deepgram Nova-3 Fast | Deepgram | 36+ | Yes | No | — | Lowest-latency Deepgram tier | — |
| AssemblyAI Universal-2 | AssemblyAI | Multilingual | Yes | No | $0.0025/min | Accurate, feature-rich transcription | — |
| AssemblyAI Universal-3 Pro ⭐ | AssemblyAI | Multilingual | Yes | No | API (pay-per-use) | New flagship STT; Universal-3 Pro + Multilingual Streaming (May 2026); Voice Agent product; LLM Gateway | — |
| Inworld AI Multilingual Streaming ⭐ | Inworld AI | Multilingual | Yes | No | API | Real-time multilingual STT for voice agents; launched May 7, 2026 alongside Universal-3 Pro | — |
| ElevenLabs Scribe v2 | ElevenLabs | 90+ | Yes | No | — | State-of-the-art transcription, word-level timestamps, diarization | — |
| ElevenLabs Scribe v2 Realtime | ElevenLabs | 90+ | Yes (~150ms) | No | — | Live transcription, ultra-low latency | — |
| Provider | Model | Cost / 1M chars | Languages | Voice Cloning |
|---|---|---|---|---|
| VibeVoice-Realtime | vibevoice-realtime-0.5b | Free | English (9 experimental) | ✅ |
| Mistral Voxtral TTS | voxtral-tts | $16 | 9 | ✅ |
| Fish Audio S2 Pro | s2-pro | $15 | 80+ | ✅ |
| OpenAI | tts-1 | $15 | 57 | ❌ |
| Grok TTS | grok-tts | $15 | 25+ | ✅ |
| Gemini 3.1 Flash TTS | gemini-3-1-flash-tts | $18.30 | 70+ | ❌ |
| Cartesia | sonic-3.5 | ~$39 | 40+ | ✅ (instant) |
| Deepgram Aura-2 | aura-2 | $30 | 7 | ❌ |
| Inworld | realtime-tts-1.5-max | $35 | 100+ | ✅ |
| ElevenLabs | flash-v2.5 | ~$60 (effective) | 32 | ✅ |
| MiniMax | speech-2.8-turbo | $60 | 40+ | ✅ |
| OpenAI | tts-1-hd | $30 | 57 | ❌ |
| Use Case | Top Picks |
|---|---|
| Real-time voice agents | Cartesia Sonic-3.5 (~100ms TTFA), Ink-2 (~100ms STT), Inworld Realtime TTS-2, Deepgram Aura-2 (~90ms) |
| Long-form narration / audiobooks | ElevenLabs v3, Gemini 3.1 Flash TTS, Fish Audio S2 Pro |
| Multilingual content | Gemini 3.1 Flash TTS (70+), ElevenLabs v3 (70+), Fish Audio S2 Pro (80+), Sonic-3.5 (40+) |
| Emotional fidelity | Hume Octave 2 (reads for meaning), ElevenLabs v3 (audio tags), Fish Audio S2 Pro |
| On-device / low cost | Kokoro-82M, MOSS-TTS-Nano, Qwen3-TTS (0.6B), Voxtral TTS (4B), VibeVoice-Realtime (0.5B) |
| Video dubbing (duration control) | IndexTTS-2 (precise duration control + emotion) |
| Enterprise / compliance | Deepgram Aura-2 (on-prem, HIPAA), Grok TTS (SOC 2 / HIPAA) |
| Open-source production | Fish Audio S2 Pro, MOSS-TTS-v1.5, IndexTTS-2, Chatterbox Multilingual v3, ZONOS2 |
Tools and frameworks for detecting unsafe content, preventing prompt injection, validating outputs, and enforcing policy compliance in LLM-powered applications. As of June 2026.
| Tool | Developer | Type | Open Source | Pricing | Best For | GitHub |
|---|---|---|---|---|---|---|
| Llama Guard 3 | Meta | Safety classifier (8B LLM) | Yes (Meta license) | Free / ~$0.02/1M tokens (API) | Input/output safety classification, 8 languages | 🔗 |
| NeMo Guardrails | NVIDIA | Programmable guardrail toolkit (Colang DSL) | Yes (Apache 2.0) | Free | Dialog safety, policy enforcement, LangChain-native | 🔗 |
| OpenAI Privacy Filter | OpenAI | PII detection & redaction | Yes (Apache 2.0) | Free (OSS) | Detects & redacts personal info in text | 🔗 |
| Guardrails AI | Guardrails AI | Python validator framework | Yes | Free (OSS) | Output validation, PII detection, hallucination guards | 🔗 |
| Amazon Bedrock Guardrails | AWS | Managed safety layer | No | Pay-per-use (AWS) | AWS-native, zero-ops compliance and content filtering | — |
| ShieldGemma 2 | Safety classifier (open weights) | Yes (open weights) | Free | Text safety (2B/9B/27B), image safety (4B) | — | |
| LLM Guard | Protect AI | Open-source middleware toolkit | Yes (MIT) | Free | PII + toxicity filtering; chains multiple scanners; drop-in middleware | 🔗 |
| Rebuff | Protect AI | Prompt injection detector | Yes | Free | Self-hardening anti-injection using vector memory | 🔗 |
| Lakera Guard | Lakera | Managed LLM security API | No | Free tier + Enterprise | Runtime LLM security, <50ms latency, PII + injection | — |
| Sponsio ⭐ | Sponsio Labs | Runtime contract enforcement | Yes (Apache 2.0) | Free (OSS) | Deterministic agent safety; trajectory-level policy enforcement; <0.01ms per tool call; OWASP Agentic Top 10 bundles; released 2026-05-06 | 🔗 |
| ToolShield ⭐ | CHATS-lab (ICML 2026) | Multi-turn safety defense | Yes (MIT) | Free | Training-free MCP tool guard; 30% attack reduction; plug-and-play for 5 coding agents | 🔗 |
| AgentGuard ⭐ | WhitzardAgent | Attribute-based access control | Yes (GPL-3.0) | Free | Tool-call-level access control; supports LangChain/AutoGen/OpenAI Agents SDK | 🔗 |
| Mirage ⭐ | ysham123 | Policy gateway | Yes (MIT) | Free | Deterministic policy DSL; same YAML for CI and production; no LLM in decision loop | 🔗 |
| ToolSafe ⭐ | MurrayTom | Step-level guardrail | Yes | Free | Proactive monitoring + feedback-driven reasoning; reduces harmful tool executions | 🔗 |
| SafeHarbor ⭐ | ljj-cyber (ICML 2026) | Memory-augmented guardrail | Yes | Free | Hierarchical Risk Tree; Safety Projector; no fine-tuning of underlying model | 🔗 |
| NEXUS ⭐ | eliashossain001 | Runtime safety monitor | Yes | Free | Structured plan IR; 4 graded interventions (allow/block/confirm/revise); deterministic + learned | 🔗 |
| mech-gov-framework ⭐ | Santander AI Lab | Mechanical governance | Yes (Apache 2.0) | Free | Model-agnostic governance regimes (R1/R2/R3); hard gates; entropy commit-reveal | 🔗 |
| Vermillio ⭐ | Vermillio | Guardrails-as-a-Service SDK | No | Paid | Real-time prompt/response filtering; copyright and data leakage protection; modular safety filters | — |
| Amazon Bedrock Guardrails InvokeGuardrailChecks API ⭐ | AWS | Per-request safeguard checks | No | Pay-per-use | Granular per-turn safety checks; numeric scores; custom thresholds; detect-only mode; released 2026-06-16 | — |
| Google AI Control Roadmap ⭐ | Google DeepMind | Agent security framework | No (internal) | Free (published) | Defense-in-depth; MITRE ATT&CK-based threat modeling; AI supervisor monitoring; D1-D4/R1-R3 levels; published 2026-06-18 | — |
Frameworks and libraries for building Retrieval-Augmented Generation (RAG) pipelines — connecting LLMs to external knowledge sources. As of June 2026.
| Framework | Developer | Language | Key Features | Open Source | GitHub |
|---|---|---|---|---|---|
| LlamaIndex | LlamaIndex | Python | 160+ data connectors, hybrid search, multi-agent support | Yes (MIT) | 🔗 |
| LangChain | LangChain AI | Python / JS | Chains, agents, memory, 50K+ integrations, LangGraph | Yes (MIT) | 🔗 |
| Dify ⭐ | LangGenius | Python / JS | 131K+ GitHub stars; visual workflow builder, RAG pipelines, MCP client/server, multi-agent, Human Input node, 1M+ apps deployed; $30M Series Pre-A (Mar 2026) | Yes (Apache 2.0) | 🔗 |
| RAGFlow | InfiniFlow | Python | Visual workflow builder, deep document parsing (PDF/tables) | Yes (Apache 2.0) | 🔗 |
| Haystack | deepset | Python | Modular pipelines, enterprise-grade, built-in monitoring | Yes (Apache 2.0) | 🔗 |
| Verba | Weaviate | Python | No-code UI, Weaviate-native vector search | Yes | 🔗 |
| Mem0 | Mem0 AI | Python / JS | Persistent memory layer, graph memory, session recall | Yes (Apache 2.0) | 🔗 |
| txtai | NeuML | Python | All-in-one semantic search + workflow automation | Yes (Apache 2.0) | 🔗 |
| R2R | SciPhi | Python | Lightweight, low-latency, REST API, production-first | Yes (MIT) | 🔗 |
| Ultra RAG ⭐ | rblake2320 | Python | 7-stage ingestion + 10-step query pipeline; KG+PPR, RAPTOR, CRAG, Self-RAG, HyDE; adversarial self-testing | Yes (MIT) | 🔗 |
| ForgeRAG ⭐ | deeplethe | Python | Production-ready; structure-aware reasoning; BM25+vector+KG; pixel-precise citations | Yes (MIT) | 🔗 |
| GRIP (ACL 2026) ⭐ | WisdomShell | Python | Retrieval-as-generation; token-level retrieval control; self-triggered information planning | Yes (MIT) | 🔗 |
| MOTHRAG ⭐ | Julian Geymonat | Python | Deterministic multi-hop RAG; research-SOTA parity on commodity APIs; no GPU; proof tree per answer | Yes (Apache 2.0) | 🔗 |
| AkasicDB / Omni RAG ⭐ | KAIST/GraphAI | SQL/GQL | Unified vector-graph-relational DBMS; Omni RAG improves accuracy 78% vs conventional RAG; 20x faster queries | Yes (SIGMOD 2026) | — |
| Ennoia ⭐ | vunone | Python | Declarative Document Indexing; typed schemas; hybrid filter+vector search; MCP tool surface | Yes (Apache 2.0) | 🔗 |
| VORTEXRAG ⭐ | vignesh2027 | Python | 7-layer pipeline solving semantic drift + context poisoning; EM=74.8 on multi-hop QA | Yes (MIT) | 🔗 |
| MemGraphRAG ⭐ | XMUDeepLIT | Python | Memory-based multi-agent graph RAG; three-layer memory architecture; KDD 2026 | Yes (MIT) | 🔗 |
| UnWeaver ⭐ | Academic | Python | Entity-based decomposition RAG; better than GraphRAG precision without explicit graph; ICLR 2026 | Yes | 🔗 |
| Google Agentic RAG ⭐ | Cloud | Multi-agent workflow with sufficient context agent; +34% accuracy vs standard RAG; public preview 2026-06-05 | No | — |
Tools and platforms for adapting pre-trained LLMs to specific tasks or domains via supervised fine-tuning, RLHF, LoRA/QLoRA, and related methods. Prices as of June 2026.
| Platform | Type | Supported Models | Pricing | Best For | GitHub |
|---|---|---|---|---|---|
| Unsloth | OSS library | Llama, Mistral, Gemma, Qwen, Phi, DeepSeek, GLM, + more | Free | 2–5× faster training, 80% VRAM reduction; MoE 12× faster (2026), FP8 RL support (1.4× faster, 60% less VRAM); Unsloth Studio web UI; Windows officially supported | 🔗 |
| Axolotl | OSS framework | Most Hugging Face models | Free | Config-as-code (YAML), reproducibility, multi-GPU training | 🔗 |
| OpenAI Fine-tuning | Managed API | GPT-4.1, GPT-4.1-mini (SFT/DPO), o4-mini (RFT) | GPT-4.1: ~$3.00/1M training tokens; GPT-4.1-mini: ~$0.80/1M | Managed, no infra; note: closed to new users as of 2026 — existing users only | — |
| Google Vertex AI | Managed cloud | Gemini 3.1 Pro/Flash, Gemma 4 | Gemini 3.1 Pro: $25/1M training tokens | GCP-native, Gemini model access | — |
| SiliconFlow | Managed cloud | 100+ open-source models | Free tier + pay-per-use | Managed fine-tuning + inference; 3-step pipeline (upload→train→deploy); 2.3× faster than avg cloud; H100/H200/MI300 | 🔗 |
| Unsloth Studio ⭐ | Unsloth | Desktop app | 500+ models | Free (OSS) / Pro / Enterprise | No-code local training and inference; runs 100% offline; GGUF/Safetensors; tool-calling; web search; released 2026-06-18 |
| Arkor ⭐ | WlyZhang | TypeScript | Open-weight LLMs | Free (alpha) | TypeScript framework for fine-tuning; type-safe configs; local Studio; managed GPUs |
| Langtrain ⭐ | Langtrain | Python | 20+ open-source models | Free (public beta) | Sovereign AI platform; fine-tune, align, deploy on own infrastructure; no per-token cost |
| Tuning Engines ⭐ | ShinyLaunch | Python | 100+ models | Free tier + paid | Unified AI control and governance layer; single OpenAI-compatible endpoint; fine-tuning + inference + guardrails |
| Predibase / LoRAX | Cloud + OSS server | Llama, Mistral, 50+ HF models | Free tier + per-GPU pricing | Multi-adapter serving: many LoRA adapters on one GPU | 🔗 |
| PEFT | Hugging Face | All Hugging Face models | Free | LoRA, QLoRA, prefix tuning, prompt tuning — full HF ecosystem | 🔗 |
| LLaMA-Factory | Community | 100+ models | Free | Web UI, low-code interface, beginner-friendly fine-tuning | 🔗 |
| torchtune | PyTorch | Llama, Gemma, Mistral, Phi | Free | PyTorch-native, composable training recipes | 🔗 |
| Pioneer (Fastino Labs) ⭐ | Managed agentic fine-tuning | Qwen, Gemma, Llama, GLiNER | API-based | First agentic fine-tuning agent; synthetic dataset generation; 10-min fine-tuning; adaptive inference | 🔗 |
| Langtrain ⭐ | Managed platform | 20+ open-source models | Public beta (free) | Sovereign AI platform; LoRA/QLoRA/DPO/GRPO; 100% on-prem; SOC 2 Type II | 🔗 |
| Fireworks Training ⭐ | Managed cloud | 100B+ models (Kimi K2.5 1T, etc.) | Preview (contact sales) | Full-parameter training; custom loss functions; frontier RL; multi-LoRA serving | 🔗 |
| Together AI Fine-tuning ⭐ | Managed cloud | 100B+ open-source models | Pay-per-use | Multi-node orchestration; 100B+ models; fine-tuning + inference in one platform | 🔗 |
| Tuning Engines ⭐ | Unified AI control layer | 100+ models | Free tier + paid | Single OpenAI-compatible endpoint; fine-tuning + routing + guardrails + policy-as-code | 🔗 |
| NeuralForge ⭐ | Local-first platform | 1-3B models on consumer GPUs | Free (OSS) | QLoRA training on consumer GPUs; web UI; GGUF export; single 12GB card sufficient | 🔗 |
| LLM Fine-Tuner v3.2 ⭐ | No-code local tool | Most HF models | Free (GPL-3.0) | No-code web UI; SFT/DPO/RLHF/ORPO; GGUF export; Unsloth 2-5x acceleration | 🔗 |
Tools for tracing LLM calls, evaluating output quality, debugging RAG pipelines, and monitoring production AI systems. Prices as of July 2026.
| Tool | Developer | Type | Open Source | Pricing | Best For | GitHub |
|---|---|---|---|---|---|---|
| LangSmith | LangChain AI | Tracing + evaluation platform | No (enterprise self-host) | Free (5K traces/mo), paid plans | LangChain apps, chain + agent debugging | — |
| Braintrust | Braintrust Data | Eval-first platform | Partial (AI proxy OSS) | Free (1M spans), enterprise | CI/CD evals, dataset management, LLM-as-judge; Topics active observability (production trace pattern discovery) GA June 2026 | — |
| Helicone | Helicone | Proxy-based observability | Yes | Free tier, usage-based | Cost tracking, request caching, drop-in API proxy | 🔗 |
| Arize Phoenix | Arize AI | OSS tracing + evaluation | Yes | Free (OSS); Arize Cloud paid | RAG debugging, LLM-as-judge, local dev | 🔗 |
| Langfuse | Langfuse | Tracing + evaluation | Yes (MIT) | Free / self-host; cloud paid | Open-source, 19K+ GitHub stars, OpenTelemetry | 🔗 |
| MLflow | Linux Foundation / Databricks | Full AI engineering platform | Yes (Apache 2.0) | Free (OSS); Databricks managed paid | 30M+ monthly downloads; observability, eval, prompt optimization, governance — no enterprise paywall | 🔗 |
| Ragas | Ragas | RAG evaluation framework | Yes | Free | RAG-specific metrics: faithfulness, recall, precision | 🔗 |
| DeepEval | Confident AI | LLM evaluation framework | Yes | Free (OSS); cloud paid | 14+ built-in metrics, pytest-style eval runner, 50+ research-backed metrics, production anomaly detection | 🔗 |
| Laminar ⭐ | Open-source observability | Yes (MIT) | Free (OSS); managed platform paid | Rust-based; ultra-fast; OpenTelemetry-native; traces + evals + AI monitoring; YC S24 | 🔗 | |
| Peekr ⭐ | Zero-config observability | Yes (MIT) | Free | Auto-instruments OpenAI/Anthropic/LiteLLM; claim-level hallucination detection; HIPAA/GDPR guardrails | 🔗 | |
| TraceMind ⭐ | Open-source eval + observability | Yes | Free | Self-hosted; automatic quality scoring; eval suites; regression alerts; hallucination detection | 🔗 | |
| Agentic CLEAR (ACL 2026) ⭐ | IBM Research | Multi-level agent eval | Yes (OSS) | Free | Automated multi-level evaluation; dynamic issue discovery; MLflow/Langfuse integration | 🔗 |
| Styxx ⭐ | Cognitive observability | Yes (MIT) | Free | 9 cognometric instruments; hallucination/refusal/tool-call drift/goal-drift detection; per-step localization | 🔗 | |
| Aether ⭐ | Local-first cognition debugger | Yes | Free | Chrome DevTools for AI; real-time reasoning tree inspection; VS Code extension; hallucination debugging | 🔗 | |
| Dynatrace dt-evals ⭐ | Dynatrace | LLM/agent eval from traces | Yes (OSS) | Free (OSS); cloud paid | Eval from real GenAI traces; LLM judge; CI/CD integration; supports OpenAI/Anthropic/Google/AWS/Azure | 🔗 |
| Currai ⭐ | Currai | AI observability platform | No | Paid | Prompt tracing; A/B testing; LLM evaluations; cost analytics; OpenTelemetry | — |
| tracesage ⭐ | Open-source | LangGraph observability | Yes (MIT) | Free | Local-first; interactive graph + timeline UI; MCP tool-source attribution; pytest fixture | 🔗 |
| Opik ⭐ | Comet | LLM lifecycle platform | Yes (Apache 2.0) | Free (OSS); cloud paid | Evaluation, testing, monitoring, optimization; Opik Guardrails | 🔗 |
| Orq.ai ⭐ | Orq.ai | Observability + monitoring | No | Paid | Real-time monitoring; automated evaluations; trace automation; custom dashboards | — |
The Model Context Protocol (MCP) is an open standard originally by Anthropic, now governed by the Agentic AI Foundation (AAIF) under the Linux Foundation (co-founded by Anthropic, Block, and OpenAI; supported by Google, Microsoft, AWS, Cloudflare, Bloomberg). It connects LLMs to external tools and data sources via a unified JSON-RPC 2.0 interface, supporting STDIO and Streamable HTTP transports. The official MCP Registry at registry.modelcontextprotocol.io is the preview centralized metadata repository for publicly accessible MCP servers, with registry docs at modelcontextprotocol.io/registry. MCP SDK downloads exceeded 97 million monthly by March 2026. The 2026-07-28 spec release candidate (locked May 21, 2026) introduces stateless core transport, Tasks extension for long-running work, MCP Apps for server-rendered UIs, OAuth 2.1 alignment, and a formal deprecation policy. Docker Custom MCP Catalogs and Profiles reached GA in May 2026. Microsoft Power BI MCP server released June 2026.
MCP Clients: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code (Copilot + ACP Client), Continue.dev, Zed, ChatGPT, Gemini, Microsoft Copilot, LibreChat, and more.
| Tool / Server | Developer | Category | Open Source | Best For | GitHub |
|---|---|---|---|---|---|
| MCP Filesystem | Anthropic / Community | File I/O | Yes (MIT) | Read/write local files from any MCP client | 🔗 |
| MCP GitHub | GitHub / Anthropic | Code & DevOps | Yes | Repo management, issues, PRs, code search | 🔗 |
| MCP Slack | Community | Messaging | Yes | Slack workspace read/write interaction | 🔗 |
| MCP PostgreSQL | Community | Database | Yes | Read-only SQL queries against Postgres | 🔗 |
| MCP Google Drive | Community | Storage | Yes | Drive file access and search | 🔗 |
| MCP Docker | Community | DevOps | Yes | Container management and inspection | 🔗 |
| MCP Brave Search | Brave | Search | Yes | Web + local search via Brave API | 🔗 |
| MCP AWS | AWS Labs | Cloud | Yes (Apache 2.0) | AWS service integration | 🔗 |
| MCP Notion | Community | Productivity | Yes | Notion page and database access | 🔗 |
| FastMCP | Community | Framework | Yes | Python framework for building MCP servers fast | 🔗 |
| Context7 | Upstash | Dev Tools | Yes | Up-to-date library docs for AI coding assistants | 🔗 |
MCP 2026-07-28 Release Candidate (May 21, 2026): The largest revision since launch. Key changes: stateless core (no sessions, no initialize handshake, plain HTTP load balancing), MCP Apps extension (server-rendered UIs), Tasks extension (long-running work), OAuth 2.0/OpenID Connect alignment, 12-month deprecation policy. Final spec ships July 28, 2026. This is a breaking change -- servers using session state must migrate to explicit handles. Source: 🔗 MCP SDK beta support (2026-06-29): Python, TypeScript, Go, and C# SDK beta releases now support the 2026-07-28 release candidate, so server maintainers can test stateless transport, Tasks, MCP Apps, and auth changes before final. Source: 🔗
Modular capability packages that extend AI agents with specialized knowledge, workflows, and procedural instructions — without bloating model context.
skills.sh is the primary registry and package manager for Agent Skills — an open standard developed by Anthropic for packaging and distributing reusable agent capabilities. Skills follow a progressive disclosure pattern: agents load only a skill's name and description at startup, then pull full instructions only when a task matches, keeping context overhead minimal.
Vercel's official Agent Resources now publish installable skills for React, Next.js, AI SDK, design/UI, browser automation, deployment, commerce, and workflow use cases. Source: 🔗
| Feature | Detail |
|---|---|
| Standard | Agent Skills (open, SKILL.md format) — developed by Anthropic, hosted on GitHub |
| Registry URL | skills.sh |
| Total installs | 90,989+ all-time |
| Compatible agents | Claude Code, Cursor, Windsurf, VS Code Copilot, Continue.dev, Zed, and any MCP-compatible agent |
| License | Open (skills are author-licensed; spec is open standard) |
| Skill | Publisher | Category | Installs |
|---|---|---|---|
| find-skills | vercel-labs/skills | Discovery | 1.3M |
| vercel-react-best-practices | vercel-labs/agent-skills | Frontend | 366K |
| frontend-design | anthropics/skills | Design | 361K |
| web-design-guidelines | vercel-labs/agent-skills | Design | 291K |
| microsoft-foundry | microsoft/azure-skills | Cloud/Azure | 286K |
| azure-ai | microsoft/azure-skills | AI/Cloud | 276K |
| agent-browser | vercel-labs/agent-browser | Browser | 229K |
| skill-creator | anthropics/skills | Meta | 180K |
| browser-use | browser-use/browser-use | Automation | 71.6K |
| systematic-debugging | obra/superpowers | Dev | 78.5K |
| test-driven-development | obra/superpowers | Dev | 68.0K |
| seo-audit | coreyhaines31/marketingskills | Marketing | 95.4K |
| supabase-postgres-best-practices | supabase/agent-skills | Database | 138K |
| playwright-best-practices | currents-dev/playwright | Testing | 34.2K |
| Publisher | Skills Count | Focus |
|---|---|---|
| microsoft/azure-skills | 19+ | Azure cloud, AI, Kubernetes, cost optimization |
| vercel-labs/agent-skills | 15+ | React, Next.js, Tailwind, deployment |
| anthropics/skills | 15+ | Design, docs, coding, web artifacts |
| coreyhaines31/marketingskills | 20+ | SEO, marketing, content, analytics |
| obra/superpowers | 12+ | Dev workflows, parallel agents, TDD |
| firebase/agent-skills | 10+ | Firebase, Firestore, GenKit |
| larksuite/cli | 13+ | Lark workspace automation |
| pbakaus/impeccable | 10+ | Design polish, code quality |
Tools for routing LLM requests across multiple providers, models, and deployments — optimizing for cost, latency, quality, or reliability. Prices as of June 2026.
| Tool | Developer | Key Features | Open Source | Pricing | GitHub |
|---|---|---|---|---|---|
| LiteLLM | BerriAI | 100+ provider support, proxy server, load balancing, fallbacks, spend tracking | Yes (MIT) | Free (OSS) / $99/mo cloud | 🔗 |
| Portkey | Portkey | 250+ LLMs, AI gateway, guardrails, observability, virtual keys | Yes (Apache 2.0) | Free tier / $49/mo+ | 🔗 |
| OpenRouter | OpenRouter | 200+ model catalog, unified API, pay-per-use credit system | No | ~5% markup on provider cost | — |
| RouteLLM | LMSys | Open-source router (strong vs. weak model) using classifier or matrix factorization | Yes | Free | 🔗 |
| Not Diamond | Not Diamond | Pre-trained + custom task-specific routers, cost/quality tradeoff | No | Free tier + enterprise | — |
| Unify AI | Unify | Quality / cost / latency-aware routing across 100+ model deployments | No | Usage-based | — |
| Semantic Router | Aurelio AI | Embedding-based semantic intent routing for agents and pipelines | Yes | Free | 🔗 |
Compact models designed for on-device inference, edge deployment, low-latency APIs, and resource-constrained environments. Generally defined as models under ~15B parameters. Specs as of June 2026.
| Model | Developer | Params | Context | License | Best For |
|---|---|---|---|---|---|
| Phi-4 | Microsoft | 14B | 16K | MIT | Reasoning, math, code — STEM benchmark leader at class size |
| Phi-4-mini | Microsoft | 3.8B | 128K | MIT | On-device STEM reasoning with long context |
| Phi-4-multimodal | Microsoft | 5.6B | 128K | MIT | Vision + audio + text multimodal, edge deployment |
| Gemma 4 31B | 31B dense | 256K | Apache 2.0 | Multimodal (text+image), hybrid-thinking, open-weight reasoning | |
| Gemma 4 26B A4B | 26B MoE (4B active) | 256K | Apache 2.0 | Multimodal, efficient MoE, agentic workflows | |
| Gemma 4 E4B | 4B dense | 256K | Apache 2.0 | On-device, CPU inference, multimodal (text+image+audio) | |
| Gemma 4 E2B | 2B dense | 256K | Apache 2.0 | Ultra-lightweight, on-device, 5GB RAM (4-bit) | |
| Gemma 4 12B | 12B dense | 256K | Apache 2.0 | Encoder-free unified multimodal (text+image+audio), hybrid-thinking | |
| Aion 1.0 Instruct ⭐ | Microsoft | — | — | Open weights (July 2026) | On-device SLM for Windows; smaller/faster than current Windows SLM; Edge Insider preview; open-source on HuggingFace July 2026 |
| Aion 1.0 Plan ⭐ | Microsoft | 14B | 32K | — | On-device reasoning/tool-calling SLM; agentic workflows locally; shipping in-box Windows "in the coming months" |
| edge-lm (Gemma 4 E2B/E4B) ⭐ | TheStageAI | 2B/4B | 256K | MIT | 7x smaller Gemma 4 checkpoints; E2B in 1.44GB; MLX-ready; optimized for Apple Silicon/edge |
| Atome LM ⭐ | TilelliLab | 60K | — | Apache 2.0 | Ternary zero-heap LM for microcontrollers; runs on ESP32; ~1 tok/s on Cortex-M3 |
| Gemma 3 27B | 27B | 128K | Apache 2.0 | Top open model, multilingual (140+ languages) | |
| Gemma 3 4B | 4B | 128K | Apache 2.0 | CPU inference, 140+ languages, mobile-friendly | |
| Gemma 3 1B | 1B | 32K | Apache 2.0 | On-device, embedded, ultra-lightweight | |
| Qwen3.5-9B | Alibaba | 9B | 256K | Apache 2.0 | Thinking + non-thinking modes; native multimodal; 201 languages; reasoning closes gap to 30B+ models |
| Qwen3.5-4B | Alibaba | 4B | 256K | Apache 2.0 | Multimodal (native VLM), agentic, on-device; edge deployment |
| Qwen3.5-2B | Alibaba | 2B | 256K | Apache 2.0 | Runs on recent iPhones (offline); 201 languages; text+image |
| Qwen3.5-0.8B | Alibaba | 0.8B | 256K | Apache 2.0 | IoT / rapid prototyping; non-thinking mode only; ultra-lightweight |
| Mistral Small 4 | Mistral AI | 119B total / 6B active (MoE, 128 experts) | 256K | Apache 2.0 | Chat + reasoning + coding + vision (multimodal) in one model; released 2026-03-16 |
| SmolLM3 | Hugging Face | 3B | 128K | Apache 2.0 | Efficient, tool use, multilingual, reasoning |
| Qwen2.5 3B | Alibaba | 3B | 128K | Apache 2.0 | Asian and multilingual tasks, coding |
| Qwen2.5 7B | Alibaba | 7B | 128K | Apache 2.0 | Strong multilingual baseline, function calling |
| Llama 3.2 3B | Meta | 3B | 128K | Llama 3.2 license | General-purpose, on-device, Meta ecosystem |
| Llama 3.2 1B | Meta | 1B | 128K | Llama 3.2 license | Lightweight edge inference, distillation target |
| Granite 3.3 8B | IBM | 8B | 128K | Apache 2.0 | Enterprise tasks, tool use, business-domain |
| MiniCPM 3.0 | ModelBest / Tsinghua | 4B | 32K | Apache 2.0 | Compact yet capable, mobile and edge |
| Danube 3 500M | H2O.ai | 500M | 8K | Apache 2.0 | Ultra-lightweight on-device, IoT |
| MobileMoE ⭐ | Academic | 0.3B–0.9B active (1.3B–5.3B total) | 128K | Research | First sub-billion MoE family for on-device; 1.8-3.8x faster than dense baselines |
| SLM-10M (Liodon AI) ⭐ | Liodon AI | 9.97M | 1K | Apache 2.0 | Leads sub-10M Open SLM Leaderboard (32.38%); beats Pythia-31M with 1/3 params |
| MiniGPT-16MB ⭐ | david-spies | 1.46M | 256 | MIT | Sub-16MB model; runs in browser via WebAssembly; 60-120+ tok/s |
| Micro Language Models (muLMs) ⭐ | Sensente | 8M–30M | — | Research | Instant on-device response initiation; masks cloud latency; collaborative generation |
| PhoneLM ⭐ | UbiquitousLearning | 0.5B / 1.5B | 32K / 128K | Apache 2.0 | Smartphone-native SLM; architecture searched for NPU efficiency; Android intent invocation |
Apple Foundation Models 3rd Generation (June 8, 2026): Apple announced five foundation models powering Apple Intelligence. AFM 3 Core (3B dense) and AFM 3 Core Advanced (20B sparse, 1-4B active) run on-device. AFM 3 Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro run on Private Cloud Compute. Not available via public API. AFM 3 Core Advanced achieves 4.15 MOS for TTS (vs 3.87 production baseline) and 44.7% preference on dictation quality. Source: 🔗
Notable GitHub repos:
Tutorials, how-tos, and in-depth guides for getting the most out of AI models and tools.
A beginner-friendly introduction to AI models and how to start using them effectively.
| Concept | Description |
|---|---|
| Parameters | Size of model (B = billions). More = more capable |
| Context Window | How much text model can process (128K standard) |
| Tokens | Basic units of text (~0.75 words per token) |
| Method | Best For | Setup Difficulty |
|---|---|---|
| Web Interfaces | Quick experiments | Easiest |
| API Access | Building applications | Easy |
| Self-Hosting | Privacy, no API costs | Medium-Hard |
| IDE Integration | Daily coding | Easy |
| Task | Free Option | Premium Option |
|---|---|---|
| Chat | Llama 4 (self-hosted) | Claude Fable 5, GPT-5.5 Instant |
| Coding | GLM-5.2, Qwen3-Coder (self-hosted) | Claude Fable 5, Claude Opus 4.8 |
| Reasoning | DeepSeek-R1 | Gemini 3 Deep Think, GPT-5.5 Pro |
| Long docs | Llama 4 Scout | Gemini 3 Flash |
| Vision | Llama 4 Maverick | GPT-5.5, Gemini 3.1 Pro |
Vibe coding — describing what you want in natural language and letting AI generate the code — has exploded in 2026. The ecosystem splits into two tracks: free AI APIs you plug into your own editor/agent, and free vibe coding IDEs/platforms that bundle everything together.
These are the raw API endpoints you can use in tools like Cursor (BYOK), Cline, or any agent framework.
| Provider | Free access | Paid baseline if you outgrow free | Best For | Source |
|---|---|---|---|---|
| Google Gemini API | Gemini Developer API free tier is free of charge | Gemini 2.5 Flash-Lite is $0.10 / $0.40 standard; $0.05 / $0.20 in Batch/Flex | Long-context and multimodal coding helpers | 🔗 |
| Groq Cloud | GroqCloud Free plan | Llama 3.1 8B Instant is $0.05 / $0.08 on paid usage | Very fast code loops and agent backends | 🔗, 🔗 |
| Cohere | Trial keys are free but limited; Command A+ is free until rate limits are reached | Command R is $0.15 / $0.60 | Tool use, RAG, and multilingual coding support | 🔗, 🔗, 🔗 |
| Mistral AI | Free mode is enabled by default with no credit card required | Mistral Small 4 is $0.15 / $0.60 on Scale | Cheap coding and multimodal experiments | 🔗, 🔗 |
| Anthropic | New users receive a small amount of free credits | Haiku 4.5 is $1.00 / $5.00 | Claude-first agent and prompt testing | 🔗, 🔗 |
| Vercel AI Gateway ⭐ | $5/month included on the free tier for eligible models | Paid usage follows provider list rates with zero markup | Multi-provider prototyping behind one endpoint | 🔗 |
| OpenAI | Developer quickstart includes one free test API request | GPT-5.4 nano is $0.20 / $1.25 | OpenAI SDK integration checks | 🔗, 🔗 |
| Tool | Type | Key Features | Best For |
|---|---|---|---|
| Cursor | AI IDE | Agent Mode, Composer 2, multi-agent workspace | Professional development |
| Cline | VS Code Extension | Open-source, BYOK/Ollama, MCP tools | Self-hosted, unlimited local LLM |
| Windsurf | AI IDE | Cascade agent, live browser preview | IDE with browser integration |
| OpenHands | Docker Agent | Self-hosted, local LLM support, full SDLC | Unlimited local development |
| bolt.diy | Browser IDE | 19+ LLM providers, Ollama, full-stack apps | Free web app building |
| Open Interpreter | CLI | Natural language → code, local LLM | Simple local automation |
Google's Chrome DevTools MCP connects AI agents directly to Chrome for debugging, profiling, and automation:
- 29 tools across 6 categories (input, navigation, emulation, performance, network, debugging)
- Run Lighthouse audits, capture performance traces, inspect network requests
- Works with Claude Code, Cursor, Copilot via MCP
- Supports BYOK/local LLMs through MCP clients
- GitHub | 37,783+ stars
Managed browser infrastructure for AI agents:
- Chrome DevTools Protocol (CDP) direct endpoint
- MCP client support (Claude, Cursor, OpenCode)
- Session recordings, Live View, WebMCP, and
/jsonschema extraction with Workers AI or BYOK - Free Workers plan / $5+/mo paid
- Browser Run
Free + Local LLM: Cline + Ollama, OpenHands + Qwen3 Coder, bolt.diy + Ollama
Fast API Iteration: Groq (speed) + Cerebras (high limits)
Web Development: Chrome DevTools MCP + Cline (zero-cost debugging)
Many Models: OpenRouter (unified API)
Production Inference: NVIDIA NIM, Cerebras
💡 Pro Tip: Combine Chrome DevTools MCP with a local LLM (Ollama) via Cline for completely free, unlimited AI-powered web development and debugging.
A comprehensive guide to running AI models on your own hardware.
| Benefit | Description |
|---|---|
| Privacy | Data never leaves your infrastructure |
| Cost Control | No per-token API costs for unlimited usage |
| Customization | Fine-tune models for specific needs |
| No Rate Limits | Process as much as hardware allows |
| Offline Access | Work without internet |
For installation and usage instructions, refer to the official Ollama documentation.
Recommended apps (local-first):
- Ollama - Simple local runtime with a local HTTP API
- LM Studio - Desktop UI for downloading and running models locally
- llama.cpp - Fast local inference (CPU/GPU), great for quantized models
- Open WebUI - Optional local web UI (pairs well with local runtimes)
If you want “server-style” hosting (advanced):
Practical setup tips:
- Install the latest NVIDIA drivers (enable GPU acceleration in your chosen app)
- Start with smaller quantized models (Q4 is a common “best default”)
- Keep context windows realistic for local hardware (lower context = faster, less memory)
- Watch VRAM first, then system RAM; reduce model size or quantization if either saturates
- Prefer running locally on
localhostand only expose to LAN if you understand firewall rules
Example hardware configurations:
| Hardware | Good starting point | Notes |
|---|---|---|
| Consumer GPU (24 GB VRAM) | 7B–14B quantized | e.g., RTX 4090, RTX 3090 — great for chat/coding |
| Pro GPU (48–80 GB VRAM) | 14B–70B quantized | e.g., A6000, A100 — coding agents, longer contexts |
| Multi-GPU (160+ GB VRAM) | 70B+ quantized | e.g., 2×A100 — larger open-source models |
| CPU-only (32–64 GB RAM) | 7B–14B quantized | Slower but viable for offline chat; keep context moderate |
| Option | Best For | Pros | Cons |
|---|---|---|---|
| Local Machine | Personal use | Simple, no latency | Limited hardware |
| Dedicated Server | Team use | Full control | Maintenance |
| Cloud GPU Rental | Experimentation | On-demand | Hourly costs |
| Kubernetes | Enterprise | Scalable | Complex |
Comprehensive pricing comparisons and cost calculations.
| Tier | Price Range | Models |
|---|---|---|
| 🆓 Free | $0 | Self-hosted, free tiers |
| 💸 Budget | $0.025 - $0.50/1M | Gemini 3.1 Flash-Lite, GPT-4.1 nano, GLM-4.7-FlashX, GPT-5.4 nano, Grok 4 Fast (aliased to 4.3) |
| 💰 Mid-range | $0.60 - $15.00/1M | GPT-5.4 mini, Claude Haiku 4.5, Kimi K2.6, Sonar, GLM-5, GPT-5.4, Claude Sonnet, Grok 4.3 |
| 💎 Premium | $15.00 - $600.00/1M | Claude Fable 5, GPT-5.5 Pro, Claude Opus 4.8, Claude Opus 4.7, o1-Pro |
AI chat apps
| Product | Plans (USD) | Notes | Official Source |
|---|---|---|---|
| ChatGPT | Go $8, Plus $20, Pro $200, Business $25/seat (annual) or $30/seat (monthly), Enterprise (contact sales) | Consumer prices are US-listed; Go is localized in some markets | 🔗 |
| Claude | Pro $20, Max $100 (5×) or $200 (20×), Team/Enterprise (see pricing) | Prices shown exclude applicable taxes; availability varies by region | 🔗 |
| Google AI (Gemini) | Free, Plus $4.99, Pro $19.99, Ultra $99.99 (5× limits / 20TB storage) or $200 (20× limits / 20TB storage) | US pricing; Plus price dropped from $7.99 on June 8, 2026 with storage doubled to 400GB; Ultra reduced from $249.99 at I/O 2026; $99.99 tier added with Gemini 3.5 Flash integration and Antigravity priority access | 🔗 |
Coding assistants
| Tool | Plans (USD) | Notes | Official Source |
|---|---|---|---|
| GitHub Copilot | Free $0, Pro $10, Pro+ $39, Business $19/user, Enterprise $39/user | Usage-based billing since June 1, 2026; AI Credits per token; free models removed; no credit rollover | 🔗 |
Comparative pricing across the cheapest official public APIs verified in this pass. Only providers and models rechecked against current official pricing pages are included below. As of 2026-07-11 19:14 UTC.
| Model | Input | Output | Cached Input | Best For | Source |
|---|---|---|---|---|---|
| Llama 3.1 8B Instant (Groq) | $0.05 | $0.08 | — | Lowest verified public direct text API price | 🔗 |
| GPT OSS 20B (Groq) | $0.075 | $0.30 | $0.0375 | Cheap open-weight routing and agents | 🔗 |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.01 | Cheapest verified multimodal API with an always-on free tier | 🔗 |
| DeepSeek-V4-Flash | $0.14 | $0.28 | $0.0028 (hit) | Strong price/performance for long-context work | 🔗 |
| Mistral Small 4 | $0.15 | $0.60 | — | Low-cost coding and multimodal work | 🔗 |
| Command R (Cohere) | $0.15 | $0.60 | — | RAG and tool use | 🔗 |
| GPT-5.4 nano | $0.20 | $1.25 | $0.02 | OpenAI lowest-cost standard model | 🔗 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 (read) | Lowest-cost Claude model | 🔗 |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.10 | Cheapest current GPT-5.6 tier | 🔗 |
| Grok 4.3 | $1.25 | $2.50 | $0.20 | xAI flagship at the current public rate card | 🔗 |
| Claude Sonnet 5 | $2.00 intro / $3.00 standard | $10.00 intro / $15.00 standard | $0.20 intro read | Strong coding model if the budget can stretch | 🔗 |
| Usage Level | Self-Host (A100) | API (GPT-5) | Winner |
|---|---|---|---|
| Light (1M tokens) | $300 (rental) | $10 | API |
| Medium (100M tokens) | $300 | $1,000 | Self-host |
| Heavy (1B tokens) | $300 | $10,000 | Self-host |
| Enterprise (10B+ tokens) | $2,000 (owned) | $100,000+ | Self-host |
Reference materials including glossary, comparison tables, and data sources.
Definitions of common terms used throughout the documentation.
| Term | Definition |
|---|---|
| Agent | AI system that autonomously performs tasks and interacts with environments |
| API | Interface for programmatically accessing AI models |
| Attention Mechanism | Neural network component focusing on relevant input parts |
| Benchmark | Standardized test measuring model performance |
| Chain-of-Thought (CoT) | Prompting technique showing step-by-step reasoning |
| Term | Definition |
|---|---|
| Fine-Tuning | Adapting pre-trained model to specific tasks |
| Frontier Model | State-of-the-art proprietary model |
| GPU | Hardware accelerator essential for ML |
| LLM | Large Language Model |
| LoRA | Efficient fine-tuning method |
| Term | Definition |
|---|---|
| MCP | Model Context Protocol for tool interaction |
| MMLU | Massive Multitask Language Understanding benchmark |
| MoE | Mixture of Experts architecture |
| Multimodal | Processing multiple input types |
| RAG | Retrieval-Augmented Generation |
| Term | Definition |
|---|---|
| Self-Hosting | Running models on own infrastructure |
| SLM | Small Language Model |
| SWE-bench | Benchmark for real GitHub issue resolution |
| Token | Basic unit of text processing |
| VRAM | GPU memory for model storage |
Side-by-side comparisons of AI models sorted by various criteria.
| 🏢 Company | 🤖 Model | 📦 Version | 📅 Release Date | 🔄 Latest Updated | 💻 Coding | 📊 Benchmarks | 💰 Price | 🖥️ Self-Host | 🔗 Official Site |
|---|---|---|---|---|---|---|---|---|---|
| 🤖 Anthropic | Claude | Fable 5 redeployed | 2026-06-09 00:00 UTC | 2026-07-01 00:00 UTC ⭐ | ✅ | SWE-bench 95.0% | $10.00 / $50.00 | ❌ | 🔗 |
| 🤖 Anthropic | Claude | Sonnet 5 | 2026-06-30 00:00 UTC | 2026-06-30 00:00 UTC ⭐ | ✅ | SWE-bench 92.4%, SWE-bench Pro 63.2%, HLE 57.4% with tools | $2.00 / $10.00 intro | ❌ | 🔗 |
| 🤖 OpenAI | GPT-5.6 | Sol / Terra / Luna | 2026-07-09 00:00 UTC | 2026-07-09 00:00 UTC ⭐ | ✅ | General availability; Sol $5/$30, Terra $2.50/$15, Luna $1/$6 | ❌ | 🔗 | |
| 🤖 OpenAI | GPT-Live | 1 / 1 mini | 2026-07-08 00:00 UTC | 2026-07-08 00:00 UTC ⭐ | — | Full-duplex voice models powering ChatGPT Voice; GPT-5.5 backend; API pending | ❌ | 🔗 | |
| 🇨🇳 MiniMax | M2 / M2.1 | Open-source | 2026-07-02 00:00 UTC | 2026-07-02 00:00 UTC ⭐ | ✅ | 230B/10B active MoE, 1M context, Apache 2.0 | ✅ | 🔗 | |
| 🌐 Google DeepMind | Gemini Omni Flash | — | 2026-06-30 00:00 UTC | 2026-06-30 00:00 UTC ⭐ | ✅ | $0.10/s video; any-to-any multimodal | ❌ | 🔗 | |
| 🌐 Google DeepMind | Nano Banana 2 Lite | — | 2026-06-30 00:00 UTC | 2026-06-30 00:00 UTC ⭐ | — | $0.034/1K images; 4s generation | ❌ | 🔗 | |
| 🇨🇳 Zhipu AI | GLM-5.2 | — | 2026-06-13 00:00 UTC | 2026-06-17 00:00 UTC ⭐ | ✅ | AIME 2026 99.2%, SWE-bench Pro 62.1%, FrontierSWE 74.4% | $1.40 / $4.40 | ✅ | 🔗 |
| 🌐 Google DeepMind | Gemini 3.5 | Flash | 2026-05-19 00:00 UTC | 2026-05-19 00:00 UTC ⭐ | ✅ | GPQA ~90.4%, SWE-bench Pro 55.1% | $1.50 / $9.00 | ❌ | 🔗 |
| 🤖 Anthropic | Claude | Mythos Preview | 2026-04-07 00:00 UTC | 2026-04-07 00:00 UTC | — | Not disclosed | $25.00 / $125.00 | ❌ | 🔗 |
| 🔬 DeepSeek | DeepSeek | V4 (Flash/Pro) | 2026-04-24 00:00 UTC | 2026-05-08 00:00 UTC | ✅ | No public benchmarks | From $0.14 / $0.28 (Flash) | ✅ | 🔗 |
| 🤖 OpenAI | GPT-5.5 | Pro | 2026-04-26 00:00 UTC | 2026-05-08 00:00 UTC | ✅ | GPQA 95.1%, SWE-bench 92.3% | $30.00 / $180.00 | ❌ | 🔗 |
| 🌐 Google DeepMind | Gemini 3 | Flash | 2026-02-12 00:00 UTC | 2026-05-08 00:00 UTC | ✅ | GPQA 90.4%, SWE-bench 78.0% | $0.50 / $3.00 | ❌ | 🔗 |
| 🚀 xAI | Grok | 4.3 | 2026-05-01 00:00 UTC | 2026-05-06 00:00 UTC | ✅ | — | $1.25 / $2.50 | ❌ | 🔗 |
| 🇨🇳 MiniMax | Hailuo | 02 | 2026-06-20 00:00 UTC | 2026-06-20 00:00 UTC ⭐ | — | #2 Video Arena | API | ❌ | 🔗 |
| 🇨🇳 MiniMax | Speech | 2.8 | 2026-01-23 00:00 UTC | 2026-07-03 21:56 UTC ⭐ | — | Native sound tags, high-fidelity cloning, studio-grade clarity | API | ❌ | 🔗 |
| 🇨🇳 MiniMax | Video-01 | — | 2026-06-20 00:00 UTC | 2026-06-20 00:00 UTC ⭐ | — | 720p/25fps | ~$0.40/video | ❌ | 🔗 |
| 🇫🇷 Mistral AI | Voxtral TTS | — | 2026-06-18 00:00 UTC | 2026-06-18 00:00 UTC ⭐ | — | 4B, 9 languages, 70ms | $0.016/1K chars | ✅ | 🔗 |
| 🤖 Microsoft | VibeVoice Realtime (0.5B) ⭐ | — | 2025-12-03 00:00 UTC | 2026-03-06 00:00 UTC ⭐ | — | 0.5B real-time TTS; <300ms TTFA; streaming input | Free (MIT) | ✅ | 🔗 |
| 🤖 Microsoft | VibeVoice ASR (7B) ⭐ | — | 2026-01-21 00:00 UTC | 2026-03-06 00:00 UTC ⭐ | — | 60-min ASR with diarization + timestamps | Free (MIT) | ✅ | 🔗 |
| 🤖 OpenAI | GPT-4.1 ⭐ | — | 2026-04-14 00:00 UTC | 2026-04-14 00:00 UTC | ✅ | 1M context, improved coding/instruction following | $2.00 / $8.00 | ❌ | 🔗 |
| 🤖 OpenAI | GPT-4.1 mini ⭐ | — | 2026-04-14 00:00 UTC | 2026-04-14 00:00 UTC | ✅ | 1M context, best value for large-context needs | $0.40 / $1.60 | ❌ | 🔗 |
| 🇺🇸 Zyphra | ZONOS2 | — | 2026-06-12 00:00 UTC | 2026-06-12 00:00 UTC ⭐ | — | 8B MoE TTS, multilingual | Free (Apache 2.0) | ✅ | 🔗 |
| 🤖 OpenAI | GPT-5.5 | — | 2026-04-26 00:00 UTC | 2026-04-26 00:00 UTC | ✅ | GPQA 93.2%, SWE-bench 88.5% | $5.00 / $30.00 | ❌ | 🔗 |
| 🤖 Anthropic | Claude | Opus 4.7 | 2026-04-22 00:00 UTC | 2026-04-26 00:00 UTC | ✅ | GPQA 94.2%, SWE-bench 87.6% | $5.00 / $25.00 | ❌ | 🔗 |
| 🏢 Company | 🤖 Model | 📅 Release Window | Notes | 🔗 Official Site |
|---|---|---|---|---|
| 🤖 OpenAI | GPT-5.6 | 2026-07 | Sol/Terra/Luna three-tier family; general availability; Sol $5/$30, Terra $2.50/$15, Luna $1/$6 | 🔗 |
| 🇨🇳 MiniMax | M2 / M2.1 | 2026-07 | Open-source MoE (230B/10B active); 1M context; Apache 2.0; $0.30/$1.20 API | 🔗 |
| 🌐 Google DeepMind | Gemini Omni Flash | 2026-06 | $0.10/s video generation; any-to-any multimodal; conversational editing | 🔗 |
| 🌐 Google DeepMind | Nano Banana 2 Lite | 2026-06 | Fastest Gemini Image; 4s gen; $0.034/1K images | 🔗 |
| 🤖 Anthropic | Claude | Fable 5 | 2026-06 | First publicly available Mythos-class model; $10/$50; free on subscription through June 22 |
| 🤖 Anthropic | Claude | Mythos 5 | 2026-06 | Restricted to Project Glasswing partners; same model as Fable 5 with safeguards lifted |
| 🇨🇳 Alibaba | Qwen3.7-Max | 2026-05 | Proprietary; $2.50 / $7.50 | 🔗 |
| 🇨🇳 Zhipu AI | GLM-5.2 | 2026-06 | 753B params, 1M context, MIT open-weight; AIME 2026 99.2%; API $1.40/$4.40 | 🔗 |
| 🌐 Google DeepMind | Gemini 3.5 Flash | 2026-05 | GA at I/O 2026; 4x faster than 3.1 Pro | |
| 🌐 Google DeepMind | Gemini Omni Flash | 2026-05 | Any-to-any multimodal creation; rolling out to AI Plus/Pro/Ultra | 🔗 |
| 🤖 Anthropic | Claude Opus 4.8 | 2026-05 | Improved Opus 4.7; SWE-bench 88.6%; fast mode $10/$50 | 🔗 |
| 🚀 xAI | Grok 4.20 / 4.3 / 4 Fast | 2026-05 | Major price cuts; Grok 4.20 now $1.25/$2.50; 4 Fast → 4.3 alias | 🔗 |
| 🤖 Anthropic | Claude Mythos Preview | 2026-04 | Invitation-only via Project Glasswing | 🔗 |
| 🤖 OpenAI | GPT-5.5 Instant | 2026-05 | Default ChatGPT model since 2026-05-05 | 🔗 |
| 🚀 xAI | Grok 4.3 | 2026-05 | Always-on reasoning; $1.25 / $2.50 | 🔗 |
| 🧠 MiniMax | MiniMax M2.5 | 2026-02 | $0.30 / $1.20 | 🔗 |
| 🇨🇳 Alibaba/Qwen | Qwen 3.5-Max | 2026-02 | Open-source release window | 🔗 |
| 🌐 Google DeepMind | Gemini 3.1 Flash-Lite | 2026-02 | Budget Gemini model | 🔗 |
| 🌐 Google DeepMind | Gemini 3 Pro | 2026-01 | Tiered pricing | 🔗 |
| 🤖 OpenAI | GPT-5.4 family | 2026-03 | GPT-5.4, GPT-5.4 mini, GPT-5.4 nano | 🔗 |
| 🤖 OpenAI | GPT-4.1 family ⭐ | 2026-04 | GPT-4.1 ($2/$8), GPT-4.1 mini ($0.40/$1.60), GPT-4.1 nano ($0.10/$0.40); 1M context; June 2024 cutoff; retired from ChatGPT Feb 13, 2026; API still available | 🔗 |
| 🇫🇷 Mistral AI | Mistral Large 3 | 2025-11 | Apache 2.0 open-source, 123B params | 🔗 |
| Rank | Model | Input | Output | License |
|---|---|---|---|---|
| 1 | Self-hosted open weights | $0 | $0 | Various |
| 2 | Llama 3.1 8B Instant (Groq) | $0.05 | $0.08 | API |
| 3 | GPT OSS 20B (Groq) | $0.075 | $0.30 | Open weights via API |
| 4 | Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Proprietary |
| 5 | DeepSeek-V4-Flash | $0.14 | $0.28 | API |
| 6 | Mistral Small 4 | $0.15 | $0.60 | Open weights/API |
| 7 | Command R | $0.15 | $0.60 | API |
| 8 | GPT-5.4 nano | $0.20 | $1.25 | Proprietary |
| 9 | Claude Haiku 4.5 | $1.00 | $5.00 | Proprietary |
| 10 | GPT-5.6 Luna | $1.00 | $6.00 | Proprietary |
| Rank | Model | SWE-bench Verified | Self-Host |
|---|---|---|---|
| 1 | Claude Mythos 5 | 95.5% | ❌ |
| 2 | Claude Fable 5 | 95.0% | ❌ |
| 3 | Claude Sonnet 5 ⭐ | 92.4% | ❌ |
| 4 | GPT-5.5 Pro | 92.3% | ❌ |
| 5 | Claude Opus 4.8 | 88.6% | ❌ |
| 6 | Claude Opus 4.7 | 87.6% | ❌ |
| 7 | GPT-5.3-Codex | 85.0% | ❌ |
| 8 | GLM-5.2 ⭐ | 62.1% (SWE-bench Pro) | ✅ |
| 9 | Claude Opus 4.6 | 80.8% | ❌ |
| 10 | Gemini 3.1 Pro | 80.6% | ❌ |
| Rank | Model | Context | Best For |
|---|---|---|---|
| 1 | Gemini 3 Flash | 10M | Entire libraries |
| 2 | Llama 4 Scout | 10M | Long-document RAG |
| 3 | Grok 4.20 | 2M | Large codebases with full context |
| 4 | Gemini 3 Pro | 1M+ | Research papers |
| 5 | Gemini 3.1 Pro | 1M | Complex multi-document analysis |
| 6 | GLM-5.2 ⭐ | 1M | Frontier open-weight coding + long context |
| 7 | Gemini 3.5 Flash | 1M | Fast multimodal reasoning |
| 8 | Claude Opus 4.7 | 1M | Agentic coding with full codebase |
| 9 | GPT-5.5 Pro | 1.05M | Premium reasoning with long context |
| 10 | Grok 4.3 | 1M | Reasoning with full project context |
Attribution, verification sources, and methodology.
| Company | Source | URL |
|---|---|---|
| OpenAI | Official Documentation | openai.com |
| OpenAI | ChatGPT agent release notes | help.openai.com |
| OpenAI | Model release notes | help.openai.com |
| OpenAI | API pricing | platform.openai.com |
| OpenAI | API pricing (overview) | openai.com/api/pricing |
| OpenAI | GPT-5.5 pro model reference | developers.openai.com |
| OpenAI | GPT-5.3-Codex model reference | developers.openai.com |
| OpenAI | March 2026 model news | openai.com |
| OpenAI | ChatGPT subscriptions (Go/Plus/Pro) | openai.com |
| OpenAI | ChatGPT Business pricing | help.openai.com |
| OpenAI | GPT-4.5 API deprecation | developers.openai.com/api/docs/deprecations |
| OpenAI | GPT-4.1 / GPT-4o retirement from ChatGPT | openai.com |
| OpenAI | GPT-Image-2 / ChatGPT Images 2.0 | openai.com |
| OpenAI | Codex changelog and CLI docs | developers.openai.com |
| OpenAI | Workspace agents in ChatGPT | openai.com |
| OpenAI | GPT-Live voice models announcement | openai.com |
| OpenAI | Realtime API (gpt-realtime 2.1) docs | platform.openai.com |
| OpenAI | Developer quickstart (free test request) | developers.openai.com |
| Anthropic | Claude Documentation | anthropic.com |
| Anthropic | Claude API pricing | docs.anthropic.com |
| Anthropic | Claude rate card and plan pricing | claude.com |
| Anthropic | Claude Sonnet 5 announcement | anthropic.com |
| Anthropic | Claude Sonnet 5 system card | anthropic.com |
| Anthropic | Claude Fable 5 redeployment | anthropic.com |
| Anthropic | Claude Opus 4.8 announcement | anthropic.com |
| Anthropic | Claude Fable 5 / Mythos 5 announcement | anthropic.com |
| Anthropic | Claude Tag announcement | anthropic.com |
| Anthropic | Claude Science workbench | anthropic.com |
| Anthropic | Claude Haiku 4.5 announcement | anthropic.com |
| Anthropic | Claude Pro pricing | anthropic.com |
| Anthropic | Max plan pricing | anthropic.com |
| Anthropic | Claude model selection and deprecations | docs.anthropic.com |
| Gemini Documentation | deepmind.google | |
| Google Cloud | Vertex AI Gemini pricing | cloud.google.com |
| Gemini Developer API pricing | ai.google.dev | |
| Gemini API models (Flash-Lite pricing) | ai.google.dev | |
| Project Mariner | deepmind.google | |
| Google AI plans | one.google.com | |
| Google AI Plus pricing | blog.google | |
| Google AI Pro pricing | one.google.com | |
| Google AI Ultra pricing | blog.google | |
| Gemini API model catalog | ai.google.dev | |
| Gemini image model cards | deepmind.google | |
| GitHub | Copilot plans & pricing | github.com |
| GitHub | Copilot changelog | github.blog |
| Kiro | IDE changelog | kiro.dev |
| Cursor | Pricing and changelog | cursor.com, cursor.com |
| Windsurf | Pricing and usage docs | windsurf.com, docs.windsurf.com |
| Zhipu AI (Z.ai) | Developer Documentation | docs.z.ai |
| Zhipu AI (Z.ai) | GLM-5.2 announcement | z.ai |
| DeepSeek | Models & API pricing; corporate site | api-docs.deepseek.com, deepseek.com |
| Mistral AI | Mistral Small 4 model card | docs.mistral.ai |
| Mistral AI | Free mode API key quickstart | docs.mistral.ai |
| MiniMax | Developer Documentation | platform.minimax.io |
| MiniMax | Pricing (Pay‑as‑you‑go) | platform.minimax.io |
| MiniMax | Speech 2.8 announcement | minimax.io |
| Moonshot AI | Developer Documentation | platform.moonshot.ai |
| Moonshot AI | Models & Pricing | platform.moonshot.ai |
| Cohere | Developer Documentation | docs.cohere.com |
| Cohere | Pricing overview and trial usage | docs.cohere.com |
| Cohere | Command R and Command A+ pricing | docs.cohere.com, docs.cohere.com |
| Groq | On-demand model pricing | groq.com |
| Groq | GroqCloud free and developer plans | groq.com |
| Cartesia | Sonic 3.5 / Ink-2 launch and docs | cartesia.ai, docs.cartesia.ai |
| AWS | Amazon Bedrock AgentCore | aws.amazon.com |
| Cloudflare | Browser Run and /json Quick Action |
developers.cloudflare.com |
| Browser Use | Browser Use pricing, cloud docs, and releases | browser-use.com, docs.browser-use.com, github.com |
| Browser Use | Browser Harness | github.com |
| Vercel | agent-browser docs | agent-browser.dev |
| Vercel | AI SDK 7 | vercel.com |
| Vercel | AI Gateway routing rules and pricing | vercel.com, vercel.com |
| Vercel | eve Agent Runs in MCP/CLI | vercel.com |
| Vercel | Sandbox FUSE support | vercel.com |
| Vercel | Agent Resources skills | vercel.com |
| ADK 2.0 and ADK Go 2.0 | adk.dev, developers.googleblog.com | |
| Genkit Agents | developers.googleblog.com | |
| Mastra | File-based agents | mastra.ai |
| DeepReinforce | Ornith-1.0 model family | huggingface.co, huggingface.co |
| Jina AI | jina-embeddings-v5-omni | jina.ai |
| Model Context Protocol | Official registry and 2026-07-28 SDK betas | registry.modelcontextprotocol.io, blog.modelcontextprotocol.io |
| AI21 Labs | Developer Documentation | docs.ai21.com |
| Perplexity | Developer Documentation | docs.perplexity.ai |
| ByteDance (Volcengine) | Developer Documentation | volcengine.com |
| Tencent (Hunyuan) | Cloud Documentation | cloud.tencent.com |
| Baidu (ERNIE) | AI Studio Documentation | ai.baidu.com |
| xAI | Grok API pricing & models | docs.x.ai, docs.x.ai |
| Apple | Foundation Models 3rd Gen announcement | machinelearning.apple.com |
| Meta | Llama Documentation | llama.meta.com |
| Benchmark | Source | Description |
|---|---|---|
| GPQA Diamond | Google Research | Graduate-level science questions (PhD difficulty) |
| MMLU-Pro | TIGER-Lab | Extended multi-task language understanding |
| Arena Elo | lmarena.ai | Crowdsourced human preference ranking |
| HLE | Scale AI | Humanity's Last Exam — expert-level questions |
| SWE-bench Verified | Princeton | Real GitHub issue resolution (human-verified) |
| SWE-bench Pro | Princeton | More challenging subset of SWE-bench |
| LiveCodeBench | LiveCodeBench | Live competitive programming problems |
| AIME 2025 | MAA | American Invitational Mathematics Examination |
| ARC-AGI-2 | ARC Prize | Abstract reasoning challenge (fluid intelligence) |
| MMMU / MMMU-Pro | MMMU | Multi-discipline multimodal understanding |
| IFEval | Google Research | Instruction-following evaluation |
| FrontierMath | Epoch AI | Expert-level research mathematics |
| HumanEval | OpenAI | 164 Python programming problems |
- Primary Source Review - Check official documentation
- Cross-Validation - Compare multiple sources
- Timestamp Verification - All data includes verification date
- Update Tracking - Monitor official channels
Additional context from the June 2026 research pass (supplements earlier tables; nothing below removes or supersedes prior entries).
- Google I/O 2026 agent announcements - Antigravity 2.0 (desktop + CLI + SDK) GA on 2026-05-19; Gemini Spark (24/7 personal agent) included with AI Ultra; Android CLI 1.0 stable for AI agent Android development; Gemini CLI sunset on 2026-06-18.
- Microsoft Build 2026 - Windows Agent Framework open-sourced (MIT); Aion 1.0 Instruct and Aion 1.0 Plan on-device SLMs announced; Windows 365 for Agents GA; Surface RTX Spark Dev Box and DGX Station for Windows announced.
- Claude Sonnet 5 launch - Anthropic launched Sonnet 5 on 2026-06-30 00:00 UTC with 1M context, 128K output, adaptive thinking, $2.00 / $10.00 introductory pricing through 2026-08-31 00:00 UTC, and 92.4% SWE-bench Verified.
- Claude Fable 5 redeployment - Anthropic restored Fable 5 and Mythos 5 access on 2026-07-01 00:00 UTC after export controls were lifted; Fable 5 is available globally, while Mythos 5 remains limited to approved Project Glasswing partners.
- Kimi K2.7 Code in GitHub Copilot - Moonshot AI's open-weight coding model became generally available in GitHub Copilot model picker on July 1, 2026; first open-weight model offered in Copilot; rolled out to Pro, Pro+, and Max plans.
- MiniMax M2 & M2.1 open-source release - MiniMax open-sourced MiniMax M2 (230B/10B active MoE, 1M context, Apache 2.0) and M2.1 (229B/10B active MoE, 1M context, Apache 2.0) on July 2, 2026; API priced at $0.30/$1.20 per 1M tokens; available on Hugging Face, vLLM, SGLang, and MiniMax Agent.
- Gemini Omni Flash and Nano Banana 2 Lite - Google released Gemini Omni Flash for video generation ($0.10/s) and Nano Banana 2 Lite for image generation (4s generation, fastest Gemini Image model) on June 30, 2026; available in Google AI Studio, Gemini API, and Gemini Enterprise Agent Platform.
- GLM-5.2 open-source release - Zhipu AI open-sourced GLM-5.2 (753B MoE, 1M context, MIT license) on 2026-06-17; available via API at $1.40/$4.40 per 1M; leads AIME 2026 at 99.2%.
- MCP 2026-07-28 spec RC - Largest revision since initial release; six breaking changes including stateless transport, two new required HTTP headers, Tasks lifecycle primitive; SDK maintainers have 10 weeks to ship support.
- GPT-Live voice models - OpenAI launched GPT-Live (1, 1 mini, Medium, High) on 2026-07-08, replacing Advanced Voice Mode inside ChatGPT; full-duplex consumer voice stack with GPT-5.5 backend delegation. No developer API at launch (Realtime API with gpt-realtime-2.1, shipped 2026-07-06, is the API path). See the TTS Proprietary table entry.
- skills.sh — Primary registry for Agent Skills packages (see Agent Skills & Registries 🎯); pairs naturally with CLI-first workflows (Claude Code, Gemini CLI, MCP-capable agents).
- CrewAI — Multi-agent framework with OSS core plus AMP Studio / AMP Cloud paths; vendor positioning emphasizes transparent orchestration, visual workflows, broad tool integrations, and team-scale adoption (see the CrewAI row in the Multi-Agent / Parallel Agent Platforms table in this document).
- Aiden — Listed above under local desktop agents as a local-first, Windows-native autonomous stack for running agents without mandatory cloud offload (taracodlabs/aiden).
Last Updated: 2026-07-11 19:14 UTC Maintained by: ReadyPixels LLC
- Awesome AI - Curated list of AI resources
- Awesome Machine Learning - ML frameworks and libraries
- Hugging Face Open Source Models - Open-source model hub
- Discussions: GitHub Discussions
- Issues: Report Issues
- 📚 Documentation: Full Docs
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Made with ❤️ by ReadyPixels LLC