[Bug]: anthropic.messages.acreate() with OpenAI target doesn't return reasoning content (uses Chat Completions instead of Responses API)

[roni-frantchi]

What happened?

When using litellm.anthropic.messages.acreate() with an OpenAI model and thinking parameter enabled, the response doesn't include actual reasoning/thinking content - only token counts.

Expected: Reasoning text returned in response (like Anthropic's thinking blocks)
Actual: Only reasoning_tokens count returned, no text content

Root cause: The anthropic adapter translates thinking → reasoning_effort and routes to Chat Completions API. But Chat Completions only returns token counts. To get actual reasoning text, OpenAI requires Responses API with reasoning: {effort: "...", summary: "auto"}.

Steps to Reproduce

import asyncio
import litellm

async def test():
    response = await litellm.anthropic.messages.acreate(
        model="openai/gpt-5.2-2025-12-11",
        messages=[{"role": "user", "content": "What is 157 * 83? Think step by step."}],
        max_tokens=4000,
        thinking={"type": "enabled", "budget_tokens": 10000},
    )
    
    content = response["content"]
    for block in content:
        print(f"Block type: {block.get('type')}")
        # Expected: thinking block with content
        # Actual: only text block, no thinking content

asyncio.run(test())

Workaround: Manually use openai/responses/ prefix and pass reasoning instead of thinking:

response = await litellm.anthropic.messages.acreate(
    model="openai/responses/gpt-5.2-2025-12-11",  # responses/ prefix
    messages=[{"role": "user", "content": "What is 157 * 83?"}],
    max_tokens=4000,
    reasoning={"effort": "high", "summary": "auto"},  # instead of thinking
)
# This returns thinking blocks with actual content

Relevant log output

# In adapters/transformation.py, thinking is translated to reasoning_effort:
reasoning_effort = translate_anthropic_thinking_to_reasoning_effort(thinking)
return {"reasoning_effort": reasoning_effort}

# Then handler.py always calls Chat Completions:
completion_response = await litellm.acompletion(**completion_kwargs)

Suggested fix

In the anthropic adapter, when targeting OpenAI with thinking enabled:

if is_openai_model and thinking:
    # Route to Responses API instead of Chat Completions
    model = model.replace("openai/", "openai/responses/", 1)
    # Use reasoning param with summary:auto to get content back
    return {"reasoning": {"effort": mapped_effort, "summary": "auto"}}

Happy to submit a PR if you're open to this approach.

Related issue

This may be related to #13419 (OpenAI gpt-5 not showing thinking outputs) - same root cause, different entry point.