Revert "make all gpt-5 and reasoning models to responses by default"

[Chesars]

Summary

PR #16766 changed mode: "chat" → mode: "responses" for multiple models (gpt-5, o1, o3, etc.). This causes severe issues:

1. Breaks old LiteLLM versions

The model_prices_and_context_window.json file is fetched dynamically from GitHub, so this change affects all LiteLLM versions, not just new ones. Users running older versions experienced unexpected breakage without updating their library.

2. Incomplete bridge implementation

The completion → responses bridge is missing response_format → text.format conversion, causing:

• JSON schemas to be silently dropped
• Structured outputs to fail
• Related: [Bug]: Sudden switch to responses from completions causing failures #16810

3. Forces unnecessary API migration

Users who don't need reasoning_content are forced to use the responses API, which has different limitations.

What this PR does

Reverts the mode field back to "chat" for models that support both endpoints:

• gpt-5, gpt-5-mini, gpt-5-nano
• o1, o1-mini, o1-preview
• o3, o3-mini, o4-mini
• etc.

Keeps mode: "responses" only for models that require it:

• o1-pro, o3-pro (only work with /v1/responses)
• gpt-5-codex, gpt-5.1-codex
• deep-research variants

No breaking changes

The supported_endpoints field already documents which endpoints each model supports. Models can work with chat/completions without losing functionality - users who want reasoning can explicitly use litellm.responses().

Next Steps

1. Implement proper response_format → text.format conversion in the bridge
2. Add intelligent endpoint selection based on request parameters
3. Address the architectural issue of mixing operation type with endpoint selection (#TBD)

Related Issues

• Fixes [Bug]: Sudden switch to responses from completions causing failures #16810

[vercel]

@Chesars is attempting to deploy a commit to the CLERKIEAI Team on Vercel.

A member of the Team first needs to authorize it.

[Chesars]

Closing as revert was made in #16849