provider: add CheapestInference

[cheapestinference]

Adds CheapestInference (https://cheapestinference.com) as a provider: flat-monthly unlimited-token subscriptions to open-weights models through an OpenAI-compatible API (https://api.cheapestinference.com/v1; an Anthropic-compatible base also exists at /anthropic).

• 6 models via base_model inheritance: kimi-k2.7 (kimi-k2.7-code), kimi-k2.6, glm-5.2, minimax-m3, deepseek-v4-flash, mimo-v2.5 — ids as served by the API
• Per-token costs are 0 following the existing subscription-plan convention — billing is a flat monthly subscription
• Monochrome currentColor logo included

Sources

• Subscription pricing (flat monthly, no per-token billing — basis for the zero per-token costs): https://cheapestinference.com/docs/pricing/plans/ — "you pay a flat monthly fee — not per-token charges"
• minimax-m3 limit.context = 1_000_000: https://cheapestinference.com/docs/models/minimax-m3/ — "Context window: 1M tokens"
• Model ids, endpoints and pool composition: https://cheapestinference.com/docs/getting-started/models/

bun run validate passes.

[github-actions]

Action items

• [medium] [possible mistake] providers/cheapestinference/models/glm-5.2.toml:2, mimo-v2.5.toml:2, minimax-m3.toml:2 - Check: reasoning_options must describe controls exposed by CheapestInference's own inference API, not the upstream model developer. Why: The proposed options are byte-identical to other providers' entries (zhipuai/glm-5.2 effort ["high","max"], xiaomi/mimo-v2.5 toggle, minimax/MiniMax-M3 toggle), yet deepseek-v4-flash.toml:2 and kimi-k2.6.toml:2 are set to [] even though deepseek/deepseek-v4-flash and moonshotai/kimi-k2.6 expose toggle/effort controls. That selective pattern, plus zero CheapestInference-specific evidence in the PR body, indicates the options were copied from upstream providers rather than verified against this provider's request surface — which the audit-reasoning-options skill explicitly prohibits. The max effort value and both toggle claims each require explicit provider evidence (the toggle claims must satisfy the " toggles reasoning with set to or " sentence). Action: For each of glm-5.2, mimo-v2.5, and minimax-m3, cite CheapestInference's API/model docs proving the exact reasoning request field and accepted values (and add the adjacent # API: {...} request-syntax comment per AGENTS.md), or set reasoning_options = [].

• [low] [possible mistake] .pr-review/pull-request.json (PR body) - Check: Data-changing PRs should cite first-party sources with each citation stating what it supports. Why: The body lists only generic doc/homepage URLs and asserts material provider-specific facts that cannot otherwise be reviewed: the context = 202_752 override for glm-5.2, context = 1_000_000 override for minimax-m3, [interleaved] field = "reasoning_content" on four models, and temperature = false on the Kimi models. None are mapped to a specific CheapestInference doc/API section. Action: Add direct links to the CheapestInference docs/API reference that justify each served-contract override (limits, interleaved field, temperature restriction), stating exactly what each link proves.

• [low] [violation] providers/cheapestinference/models/kimi-k2.6.toml:3, providers/cheapestinference/provider.toml:3 - Check: In-file source/rationale comments must live in a leading comment block above the first key. Why: AGENTS.md requires this so sync serialization does not silently delete comments; the # The upstream coding-plan API rejects temperature values other than 1. comment sits between reasoning_options and temperature, and the subscription-pricing rationale in provider.toml sits between npm and doc. Action: Move each rationale comment block to the top of its file, above the first key.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [medium] [possible mistake] providers/cheapestinference/models/deepseek-v4-flash.toml:1 - Check: interleaved field for a reasoning model served via an OpenAI-compatible API that returns reasoning_content. Why: interleaved is provider-specific (the ModelMetadata schema excludes it, so it cannot be inherited from base_model) and must be declared in the provider TOML when the API returns interleaved reasoning. Every other model in this PR (glm-5.2, kimi-k2.6, kimi-k2.7, mimo-v2.5, minimax-m3) declares [interleaved] field = "reasoning_content", and all 21 existing provider entries for deepseek-v4-flash across the repo declare the same. The PR body states the API was probed for deepseek-v4-flash (it is listed among the models checked for reasoning_content), so the served API almost certainly returns reasoning_content for this model too. The omission leaves consumers without the signal that reasoning content is interleaved. Action: Verify whether CheapestInference's API returns reasoning_content for deepseek-v4-flash; if it does (as the repo-wide pattern and the other five models in this PR indicate), add [interleaved] with field = "reasoning_content".

[github-actions]

Action items

• [low] [possible mistake] providers/cheapestinference/models/deepseek-v4-flash.toml - Check: Interleaved reasoning field consistency across the new provider's reasoning models. Why: The PR body states reasoning_content was verified against the served API as the interleaved field, and 5 of the 6 new models declare [interleaved] field = "reasoning_content"; deepseek-v4-flash is the only reasoning model in this PR without it. DeepSeek V4 Flash is a reasoning model that every other provider in the repo serves with interleaved.field = "reasoning_content", so the omission looks like an oversight rather than a provider-specific difference. Action: Add [interleaved]\nfield = "reasoning_content" to deepseek-v4-flash.toml if the served API returns reasoning_content, or confirm in the PR body that this model specifically does not expose interleaved reasoning on CheapestInference.
• [low] [possible mistake] providers/cheapestinference/models/kimi-k2.6.toml:2 - Check: Evidence backing reasoning_options = [] on the five non-toggle reasoning models. Why: The PR documents concrete API tests only for mimo-v2.5 (toggle found) and kimi-k2.6 (temperature 400), then asserts "no request-level reasoning control is exposed" for the remaining reasoning models without describing which reasoning fields were probed. kimi-k2.6 is a hybrid model that exposes an enable_thinking toggle on other providers, so the negative claim is material and cannot be reviewed from the diff alone. Action: State in the PR body which reasoning request fields (e.g. enable_thinking, reasoning_effort, chat_template_kwargs) were tested against the served API for the reasoning_options = [] models and the observed behavior (silently ignored vs. rejected), so the empty arrays are substantiated per the audit-reasoning-options evidence standard.

[github-actions]

Action items

• [medium] [possible mistake] providers/cheapestinference/models/mimo-v2.5.toml:4 - Check: Reasoning-option toggle claims must be backed by inference-provider API evidence. Why: mimo-v2.5 declares a toggle (API: {"chat_template_kwargs": {"enable_thinking": false}}), but the PR body cites no CheapestInference documentation proving its OpenAI-compatible gateway accepts/forwards that field for this model, while the other five models correctly use reasoning_options = [] for unverified controls. Upstream MiMo/vLLM serving syntax does not prove a gateway exposes the control, and the repo standard states an empty array is correct when no caller control is verified. Action: Cite CheapestInference API/model documentation that documents chat_template_kwargs.enable_thinking for mimo-v2.5, or change reasoning_options to [].

[github-actions]

Action items

• [high] [possible mistake] providers/cheapestinference/models/glm-5.2.toml - Check: Provider-specific limit overrides must be applied when the provider serves a different context window than the base model. Why: The PR body explicitly cites limit.context = 202_752 for glm-5.2 ("Context window: 198K tokens") with a source URL, but the committed TOML has no [limit] block and inherits limit.context = 1_000_000 from models/zhipuai/glm-5.2.toml. The catalog will therefore publish a 1M context window for this provider, directly contradicting the PR's own cited evidence. The same PR correctly applies the analogous override for minimax-m3, so this is an omission, not an intentional choice. Action: Add [limit] context = 202_752 (and an output override if the provider's output cap differs) to glm-5.2.toml, or correct the citation if 1M is actually served.
• [medium] [possible mistake] providers/cheapestinference/models/mimo-v2.5.toml:3 - Check: reasoning_options toggle claims require provider-specific evidence that the same model ID toggles reasoning via the documented request, not just upstream chat-template capability. Why: This is the only model in the PR that declares a toggle (# API: {"chat_template_kwargs": {"enable_thinking": false}}); the other five reasoning models use reasoning_options = []. The PR body cites only model IDs/endpoints and context windows — nothing documents that CheapestInference, an OpenAI-compatible gateway, accepts or forwards chat_template_kwargs.enable_thinking for mimo-v2.5. The audit-reasoning-options skill prohibits inferring a toggle from the upstream model's chat template and requires provider documentation or reproduction. Action: Provide a provider citation or reproduction showing CheapestInference accepts that request field for mimo-v2.5, or change to reasoning_options = [] for consistency with the other models.
• [low] [possible mistake] providers/cheapestinference/models/deepseek-v4-flash.toml - Check: interleaved should be declared when the provider streams interleaved reasoning_content, consistent with the PR's other reasoning models. Why: Five of the six new reasoning models in this PR declare [interleaved] field = "reasoning_content"; deepseek-v4-flash is the only one without it, and 22 other providers serving deepseek-v4-flash declare the same interleaved field. The within-PR inconsistency suggests an oversight. Action: Confirm whether CheapestInference streams interleaved reasoning_content for deepseek-v4-flash and add the [interleaved] block if so.
• [low] [possible mistake] providers/cheapestinference/provider.toml (PR body) - Check: Material factual pricing changes should cite a direct provider pricing/subscription source. Why: The cost = 0 flat-monthly subscription convention is a material factual claim that cannot be reviewed from the included evidence; the PR body links to model docs but not a pricing/subscription page that supports zero per-token billing. Action: Add a citation to the provider's pricing or subscription page that supports the zero per-token costs.

[cheapestinference]

All four items addressed in the latest push:

• glm-5.2: resolved — the [limit] override and its citation were removed in the current revision; the base model card's window applies to this entry.
• mimo-v2.5: changed to reasoning_options = [], consistent with the other models.
• deepseek-v4-flash: [interleaved] field = "reasoning_content" added, consistent with the other entries.
• Pricing: added a citation to the subscription pricing page in the PR body supporting the zero per-token costs.

bun run validate passes.

[github-actions]

Action items

• [low] [possible mistake] providers/cheapestinference/provider.toml:1 - Check: Cite a first-party source for the flat-monthly subscription / zero per-token billing claim. Why: Setting cost.input = 0, cost.output = 0, and cost.cache_read = 0 across all six new model files is a material pricing assertion. The PR body's two citations cover only the minimax-m3 context window and the model IDs/endpoints; neither documents the subscription billing model, so the zero-cost data cannot be reviewed from the provided evidence. Action: Add a direct link to CheapestInference's pricing or subscription-plan page in the PR body, stating that it supports the flat-monthly unlimited-token billing model that justifies the 0 per-token costs.

[cheapestinference]

pls check: Subscription pricing (flat monthly, no per-token billing — basis for the zero per-token costs): https://cheapestinference.com/docs/pricing/plans/ — "you pay a flat monthly fee — not per-token charges"

[github-actions]

Action items

• [low] [possible mistake] providers/cheapestinference/models/glm-5.2.toml ([limit] context = 202_752) - Check: Provider-specific limit.context overrides must be supported by a provider source. Why: This file overrides the inherited GLM-5.2 context window (1,000,000 from models/zhipuai/glm-5.2.toml) down to 202,752 — a ~5x reduction introduced in the follow-up "fix" commit. The PR body cites a source for minimax-m3's 1M context but gives no source for GLM-5.2's 202,752 value, and the TOML has no leading citation comment. The non-round value cannot be reviewed without evidence, and an incorrect override would misstate the served context window in the generated catalog. Action: Add a citation to the provider's GLM-5.2 model page/API confirming the 202,752 context window (as a leading TOML comment and in the PR body), or remove the [limit] override to inherit the upstream 1M context.
• [low] [possible mistake] providers/cheapestinference/models/*.toml ([interleaved]) - Check: [interleaved] should be declared when the provider returns reasoning content in reasoning_content. Why: The initial provider-authored commit declared [interleaved] field = "reasoning_content" on all six reasoning models, then the follow-up "fix" commit removed it from all six with no explanation in the PR body. For OpenAI-compatible serving of these reasoning models, reasoning_content is the standard interleaved field, and peer subscription providers of the same base models (e.g. zai-coding-plan, umans-ai-coding-plan, minimax-coding-plan) declare it. The diff's own history is contradictory evidence that the provider returns interleaved reasoning, so the final omission may drop accurate metadata. Action: Confirm whether the OpenAI-compatible API returns reasoning in reasoning_content; if so, re-add [interleaved] field = "reasoning_content" to the reasoning model files.

[github-actions]

I have all the context I need. Let me reconstruct the final state of the changed files and compare against the schema, base-model metadata, and the reasoning-options audit standard, including how comparable subscription providers serving the same base models document their reasoning controls.

Action items

• [medium] [possible mistake] providers/cheapestinference/models/{deepseek-v4-flash,glm-5.2,kimi-k2.6,kimi-k2.7,mimo-v2.5,minimax-m3}.toml:3 - Check: reasoning_options must reflect CheapestInference's actual API surface, not an assumed OpenAI-compatible baseline. Why: Patch 3 replaces the originally-correct reasoning_options = [] with [[reasoning_options]] type = "effort" values = ["low","medium","high"] on all six reasoning models, but the PR body cites no provider API reference or model doc proving CheapestInference accepts a reasoning_effort-style control with those values. The audit-reasoning-options skill states that for OpenAI-compatible APIs low/medium/high are "a useful investigation baseline, not proof," and that each value must be verified separately with provider evidence. The commit message ("add standard reasoning efforts") indicates the values were added by assumption. Comparable subscription providers serving the same base models document different, provider-verified controls — e.g. umans-glm-5.2 and alibaba-token-plan-cn/glm-5.2 expose effort values ["high","max"] plus a toggle; umans-kimi-k2.7 exposes [] (reasoning always on); alibaba-token-plan-cn kimi-k2.6, deepseek-v4-flash, and minimax-coding-plan/MiniMax-M3 expose a toggle, not low/medium/high effort — which makes the blanket low/medium/high claim suspicious. Action: Cite CheapestInference's API reference or model documentation showing, per model, the exact reasoning request field and accepted values it honors; otherwise revert to reasoning_options = [] for any model where no caller-facing control is verified.
• [low] [possible mistake] providers/cheapestinference/models/glm-5.2.toml:13 - Check: The [limit] context = 202_752 override needs a source. Why: This overrides the inherited zhipuai/glm-5.2 context window (1,000,000) down to an unusual 202,752, a material factual change that the PR body does not cite or explain (it cites the 1M window only for minimax-m3). The value cannot be reviewed without a provider reference. Action: Add a citation to CheapestInference's model docs/API showing the served context window for glm-5.2, or drop the override so it inherits the base model's limit.