GPT-5.5 Pro fails with 404 because it is routed to Chat Completions instead of the Responses API

[cdamus]

Bug Description:

The official OpenAI provider ships several -pro models in its default model list (gpt-5.5-pro, gpt-5.4-pro), but selecting one fails with:

404 This is not a chat model and thus not supported in the v1/chat/completions endpoint. Did you mean to use v1/completions?

-pro models (like OpenAI's gpt-5-pro/o3-pro) are served only by the Responses API (v1/responses); they are not available on v1/chat/completions. Theia chooses the endpoint solely from the global preference ai-features.openAiOfficial.useResponseApi, which defaults to false. There is no per-model routing, so a -pro model is sent to Chat Completions and 404s out of the box.

Plain gpt-5.5 works because it is served on both endpoints; only the -pro variants are affected.

Expected: A model that is shipped in the default list works when selected, without requiring the user to manually flip a global preference.

Related: When useResponseApi is enabled but a Responses-API call fails, OpenAiModel.handleResponseApiRequest silently falls back to Chat Completions, which then surfaces this same 404 and masks the real underlying error. That fallback makes the root cause harder to diagnose.

Possible fix: Resolve the endpoint per model (e.g. add a useResponseApi/responsesOnly flag to the -pro family in openai-model-defaults.ts) so Responses-API-only models self-route regardless of the global preference.

Workaround: Set the preference ai-features.openAiOfficial.useResponseApi to true. Note this routes all official OpenAI models through the Responses API.

Steps to Reproduce:

1. Configure the official OpenAI provider with a valid API key, leaving ai-features.openAiOfficial.useResponseApi at its default (false).
2. Open the AI chat and select the gpt-5.5-pro model.
3. Send any message.

The request fails with the 404 ... not supported in the v1/chat/completions endpoint error above.

Additional Information

• Operating System: macOS
• Theia Version: 1.73.0

[ndoschek]

Thanks for the report @cdamus!
@eneufeld could you have a quick loook?

[eneufeld]

@ndoschek I can check tomorrow. But basically we should go away from the completions api and only use the response api. or use response api by default and not based on some preference.