gpt-5.6-luna advertised in /models but not accessible via /chat/completions (breaks MoA/aggregator tooling)

[AbhiGullz]

Summary

gpt-5.6-luna is advertised as available in the Copilot Models API (GET /models), but it cannot be used via the /chat/completions endpoint — the standard OpenAI-compatible surface. It only works through /responses, which breaks tooling that relies on chat completions (e.g. Mixture-of-Agents workflows, aggregators, OpenAI-compatible SDK clients).

Environment

• Endpoint: https://api.githubcopilot.com
• Auth: GitHub Copilot individual subscription (token from gh auth token)
• Integrator: copilot-language-server
• Headers used: Editor-Version: vscode/1.104.1, Openai-Intent: conversation-edits, x-initiator: agent

Steps to reproduce

1. GET https://api.githubcopilot.com/models → response includes:

   {"id": "gpt-5.6-luna", "capabilities": {"limits": {...}}}

   (also listed in GET https://api.githubcopilot.com/models?intent=chat)

2. POST https://api.githubcopilot.com/chat/completions with:

   {"model": "gpt-5.6-luna", "messages": [{"role": "user", "content": "Say OK"}], "stream": true}

   → HTTP 400:

   {"error": {"message": "model \"gpt-5.6-luna\" is not accessible via the /chat/completions endpoint", "code": "model_not_supported", "type": "invalid_request_error"}}

3. The same request to POST https://api.githubcopilot.com/responses works (returns a valid response).

Expected behavior

A model that is advertised in GET /models should be callable through the standard /chat/completions endpoint. If a model is restricted to a specific API surface, the /models response should expose that restriction (e.g. a supported_endpoints / api_modes field), so clients can discover it without trial-and-error.

Impact

• Breaks any client that uses the OpenAI-compatible chat completions interface (most third-party tools, aggregators, Mixture-of-Agents runtimes).
• Confusing developer experience: model appears in the picker/list but fails at runtime with a non-obvious error.
• No way to discover the endpoint restriction from the API itself — only by hitting the error.

Notes

• Same class of issue as Execution failed: CAPIError: 400 model "gpt-5.3-codex" is not accessible via the /chat/completions endpoint #2177 (gpt-5.3-codex not accessible via /chat/completions) — likely affects other GPT-5.x models too.
• Happy to provide more details or test additional models if useful.

[andyfeller]

First, thank you for the report and apologies for the frustration this might have caused, @AbhiGullz! 🙇

I've been able to confirm this and a few other observations that I'm going to work internally to understand and hopefully resolve.