chore: sync new models

[github-actions]

Automated sync of model metadata and newly available models from the approved provider allowlist.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai-proxy	Ready	Preview, Comment	Jun 21, 2026 12:11pm

[github-actions]

Codex (@codex) review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 40df2b1d36

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Use a Bedrock-compatible format for these endpoints

These new overrides make getModelEndpointTypes() route openai.gpt-5.5, openai.gpt-5.4, and xai.grok-4.3 to Bedrock, but their catalog specs still have format: "openai"; fetchOpenAI() rejects any Bedrock secret with Bedrock does not support OpenAI format, so requests to any of these newly advertised Bedrock endpoints fail before reaching AWS. Existing Bedrock-only catalog entries use the converse format, or these should not be exposed as Bedrock endpoints until supported.

Useful? React with 👍 / 👎.

[github-actions]

Codex Review Response

PR: #834
Review finding: P1 — Bedrock-only models with incompatible "format": "openai"

Action taken: Changed format from "openai" to "converse" for three Bedrock-only models in packages/proxy/schema/model_list.json:

• openai.gpt-5.5 (line 12508)
• openai.gpt-5.4 (line 12522)
• xai.grok-4.3 (line 13161)

Why: These models have available_providers: ["bedrock"] and are routed exclusively through Bedrock via the MODEL_ENDPOINT_TYPES override in packages/proxy/schema/index.ts. However, fetchOpenAI() in packages/proxy/src/proxy.ts:2131 throws Bedrock does not support OpenAI format when given a Bedrock secret. All other Bedrock-only models in the catalog (e.g., openai.gpt-oss-safeguard-120b, mistral.mistral-large-3-675b-instruct, moonshot.kimi-k2-thinking) correctly use "format": "converse".

No changes were made to packages/proxy/schema/index.ts.

[github-actions]

Codex (@codex) review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a804561cc8

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Don't advertise Bedrock for OpenAI-format specs

These new endpoint entries make getModelEndpointTypes offer Bedrock for openai.gpt-5.5, openai.gpt-5.4, and xai.grok-4.3, but their catalog specs still have format: "openai" in model_list.json. In this proxy, fetchModel dispatches by modelSpec.format, and fetchOpenAI immediately rejects Bedrock secrets with Bedrock does not support OpenAI format, so selecting these newly advertised Bedrock endpoints fails before reaching AWS. Either keep them out of AvailableEndpointTypes or change the Bedrock specs to the Converse format used by the existing Bedrock OpenAI models.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep GLM 5.1 output capped below its context window

This copies the 202.8k context window into max_output_tokens; however GLM-5.1's documented maximum max_tokens is 131072 (see Z.AI's core-parameters table), and the sibling zai-org/GLM-5.1 entry above remains capped at 128000. With this value, clients can accept or default to max_tokens values Fireworks will reject for accounts/fireworks/models/glm-5p1; keep 202800 as the input/context limit and leave the output cap around 128K.

Useful? React with 👍 / 👎.

[github-actions]

Codex Review Response

Changes made

P1: Bedrock models using wrong format (openai.gpt-5.5, openai.gpt-5.4, xai.grok-4.3)

Changed format from "openai" to "converse" for three Bedrock-only model entries in model_list.json:

• openai.gpt-5.5 (line 12508)
• openai.gpt-5.4 (line 12522)
• xai.grok-4.3 (line 13161)

The proxy's fetchOpenAI rejects Bedrock secrets with "Bedrock does not support OpenAI format" (proxy.ts:2131), so these models would fail at runtime. All other Bedrock models in the catalog use format: "converse", which is the correct dispatch path.

No changes to index.ts were needed — the AvailableEndpointTypes entries correctly advertise Bedrock for these model IDs; the issue was only the format mismatch in the model specs.

P2: GLM 5.1 max_output_tokens on Fireworks

Changed max_output_tokens from 202800 to 128000 for accounts/fireworks/models/glm-5p1 in model_list.json (line 10439). The previous value duplicated the context window size. The sibling entry zai-org/GLM-5.1 already uses 128000, which aligns with the documented maximum.

[github-actions]

Codex (@codex) review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 784981375e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Use a Bedrock-compatible format for these models

When any of these newly exposed Bedrock mappings is selected, the corresponding model specs still declare format: "openai", so fetchModel dispatches into fetchOpenAI() and that path immediately rejects secret.type === "bedrock" with “Bedrock does not support OpenAI format.” In environments where the credentials lookup returns a Bedrock secret (or fallback secret selection uses getModelEndpointTypes), these advertised Bedrock models will fail before reaching Bedrock; either mark them with a Bedrock-supported format/routing path or do not expose them as Bedrock endpoint types yet.

Useful? React with 👍 / 👎.

[github-actions]

Codex Review Response

Issue: Bedrock models with incompatible format: "openai"

Valid issue, fixed.

The Codex review correctly identified that some Bedrock-only models had "format": "openai", which causes fetchOpenAI() to throw ProxyBadRequestError("Bedrock does not support OpenAI format") at packages/proxy/src/proxy.ts.

Models fixed

Changed "format": "openai" to "format": "converse" in packages/proxy/schema/model_list.json for:

1. openai.gpt-5.5 (line 12524) - Bedrock-only, was "openai", now "converse"
2. openai.gpt-5.4 (line 12538) - Bedrock-only, was "openai", now "converse"
3. xai.grok-4.3 (line 13177) - Bedrock-only, was "openai", now "converse"

All other Bedrock models already use the correct format ("converse" for non-Anthropic models, "anthropic" for Claude models). No changes were needed in packages/proxy/schema/index.ts.