I think I'm holding it wrong? Your fork/implementation of Codex explains that we can use Anthropic models.

[shaneholloman]

However, when I open up Code using the --model flag and pick claude-opus, Claude Opus is successfully picked, but it will not run.

 ✖ unexpected status 400 Bad Request: {"detail":"The 'claude-opus' model is not supported when using Codex with a ChatGPT account."}

I would imagine either I'm doing something wrong or this is not being implemented yet such that we can use freely different models?

Sorry if this is the dumbest question you've had all day!

[ConstantinHvber]

You could also just use https://github.com/musistudio/llms to translate between the Claude Messages API and the Responses V1 Api Codex/Code uses. Its a workaround but it works quite nicely

[shaneholloman]

@ConstantinHvber very cool, thanks for the heads up. Is that the official solution for this project, or is this something that you think would be useful to add to 'Code'?

I'm asking this because it's advertising multiple providers, yet I can't make that work.

[ConstantinHvber]

@shaneholloman https://github.com/just-every/code/blob/main/docs/agents.md
So the multiple providers currently relates only to the agents that are created within the main session that uses GPT. Their own API is not implemented but rather they are called directly with their providers CLI in headless mode.

However you can use Claude etc via https://github.com/musistudio/llms by using it to run a simple proxy in the background that converts the Responses V1 calls that code does to Messages and then forwards them to the Claude API. Then you'd just need to set the OPENAI_BASE_URL env var to your proxy. That should be vibecodable in half an hour or something, I did something similar to use my Codex Account within Claude Code

[shaneholloman]

@ConstantinHvber ... ahhh (KLUNK!) me 'gets' it now, total sense! cheers mate ...

[ConstantinHvber]

Still your issue makes sense
--model <name> Override the model (gpt-5.1, claude-opus, etc.)
in the README does have to be fixed and does imply what you were thinking earlier.

[shaneholloman]

@ConstantinHvber I appreciate you noticing that. That was my driver for presenting the issue today. AS it seemed like is was planned at one point

[zemaj]

Thanks for raising this — you weren’t doing anything wrong. --model only changes the model name sent to the active provider; it doesn’t switch providers. The primary session expects an OpenAI‑compatible HTTP API (Chat Completions or Responses). Native Anthropic/Gemini APIs aren’t supported directly unless you put an OpenAI‑compatible proxy in front.

I’ve updated the docs to clarify this and added a proxy example to the config docs. If you want Claude Opus as the primary model today, point model_provider at a Responses/chat‑compatible proxy and set wire_api accordingly.

[zemaj]

Closing since this is expected behavior; docs updated to clarify OpenAI‑compatible providers + proxy guidance. Reopen if you’d like native provider support tracked as a feature request.