GPT-5.6 Luna returns model not found with ChatGPT OAuth

[AidenGeunGeun]

Description

gpt-5.6-luna is listed under the built-in OpenAI provider with ChatGPT OAuth, but requests fail with HTTP 404:

Model not found gpt-5.6-luna

I reproduced this from a clean anomalyco/opencode dev checkout at c2e019ba7. The same ChatGPT account succeeds with gpt-5.6-luna through Codex CLI 0.144.0.

Codex 0.144 marks Luna with minimal_client_version: 0.144.0 and use_responses_lite: true. OpenCode currently admits the model through its > 5.4 OAuth filter but sends its normal Responses request path using the OpenCode installation version. This looks like an OpenCode Codex-transport compatibility gap rather than a bad model slug.

This is related to, but narrower than, #36133: that report includes custom proxy/API-key configuration and the gpt-5.6 alias. This reproduction uses only the built-in ChatGPT OAuth path and the exact gpt-5.6-luna slug.

A focused implementation is available in #36143.

Plugins

None (--pure).

OpenCode version

0.0.0-dev-202607091906, built from upstream dev commit c2e019ba7 (v1.17.18 source).

Steps to reproduce

1. Build a clean upstream checkout with OPENCODE_CHANNEL=dev bun run script/build.ts --single from packages/opencode.
2. Authenticate the OpenAI provider using ChatGPT OAuth.
3. Run opencode run --pure --model openai/gpt-5.6-luna "Reply with exactly: LUNA_OK".
4. Observe repeated provider retries; the endpoint response is HTTP 404 with Model not found gpt-5.6-luna.

Expected: Luna responds, as it does through Codex CLI 0.144.0 using the same account.

Screenshot and/or share link

No share link; this was reproduced in an isolated development database. The exact command and backend error are above.

Operating System

macOS 26.6, Apple Silicon.

Terminal

Ghostty

[AidenGeunGeun]

I’d like to work on this. My proposed first step is to mirror Codex 0.144’s GPT-5.6 request contract in the OpenAI OAuth adapter and cover it with a credential-free local-server regression test. Please let me know if this is already underway or if you prefer a different boundary.

EDIT: Pr available at #36143

[adamrybak]

Same issue for me although sol and terra are working fine.

[janvitos]

Same issue here.

Model not found gpt-5.6-luna [retrying in 2s attempt #2]

[emcd]

Same issue for me although sol and terra are working fine.

I had a similar problem with gpt-5.6-sol in the earlier issue which was closed as a duplicate of this one: #36133 (comment)

That said, I just tried it again (6 hours later, same Opencode version) and it works now. Not sure if there something provider-side at play here. Possibly Openai has relaxed the request contract since initial release. So, maybe it is actually specific to Luna.

[Eavheang]

I got the same issue with GPT-5.6-Luna model

RequestExecutor.execute: Provider request failed with HTTP 404: {
  "error": {
    "message": "Model not found gpt-5.6-luna",
    "type": "invalid_request_error",
    "param": "model",
    "code": null
  }
}

Although Terra and Sol work fine for me.

[narudesu]

I can confirm, Sol + Terra work ok for me, while Luna presents with the Model not found gpt-5.6-luna [retrying in xs attempt #x] error.

In the official codex CLI Luna works ok for me.

[henri318]

same issue sol and terra are working, codex is working to

[Fanduzi]

same issue

[profitevillager]

same

[ManuelZierlSubstain]

probably same problem as here: vectorize-io/hindsight#2643
so this might be worth a look: vectorize-io/hindsight#2647