What happened?
When using litellm.aanthropic_messages() with a non-Anthropic model (e.g. openai/gpt-5.1) and passing thinking={"type": "enabled", "budget_tokens": 5000, "summary": "concise"}, the summary field is silently dropped and replaced with a hardcoded "detailed".
Expected: The user's summary: "concise" should be preserved and sent to OpenAI's Responses API.
Actual: summary is dropped by translate_anthropic_thinking_to_reasoning_effort() (returns plain string "medium"), then _route_openai_thinking_to_responses_api_if_needed() hardcodes summary: "detailed".
Steps to Reproduce
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import LiteLLMAnthropicMessagesAdapter
from litellm.llms.anthropic.experimental_pass_through.adapters.handler import LiteLLMMessagesToCompletionTransformationHandler
adapter = LiteLLMAnthropicMessagesAdapter()
# User passes summary="concise"
thinking = {"type": "enabled", "budget_tokens": 5000, "summary": "concise"}
# Step 1: translate_anthropic_thinking_to_reasoning_effort drops summary
effort = adapter.translate_anthropic_thinking_to_reasoning_effort(thinking)
print(f"reasoning_effort: {effort!r}") # 'medium' — summary is gone
# Step 2: translate_thinking_for_model wraps it
result = adapter.translate_thinking_for_model(thinking, "openai/gpt-5.1")
print(f"result: {result!r}") # {'reasoning_effort': 'medium'} — no summary
# Step 3: handler hardcodes summary="detailed"
completion_kwargs = {"model": "openai/gpt-5.1", "reasoning_effort": "medium"}
LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed(
completion_kwargs, thinking=thinking
)
print(f"final: {completion_kwargs!r}")
# {'model': 'responses/openai/gpt-5.1', 'reasoning_effort': {'effort': 'medium', 'summary': 'detailed'}}
# Expected summary: "concise", got: "detailed"Suggested Fix
The fix is small — translate_anthropic_thinking_to_reasoning_effort() should return a dict instead of a string when summary is present in the thinking config:
# In LiteLLMAnthropicMessagesAdapter.translate_anthropic_thinking_to_reasoning_effort():
#
# Instead of always returning a plain string like "medium",
# return {"effort": "medium", "summary": "concise"} when summary is present.
# This way _route_openai_thinking_to_responses_api_if_needed() will see it's
# already a dict with "summary" and won't hardcode "detailed".
summary = thinking.get("summary")
if summary:
return {"effort": effort_str, "summary": summary}
return effort_strThe handler already handles dicts correctly (lines 76-83 of handler.py) — it only adds "summary": "detailed" when the key is absent.
We'd be happy to open a PR for this if you'd accept one.
Workaround: We currently strip thinking from the kwargs and pass reasoning_effort as a dict directly to bypass the translation.
Relevant log output
Step 1 - reasoning_effort result: 'medium'
-> summary field is LOST (returned plain string, not dict)
Step 2 - translate_thinking_for_model result: {'reasoning_effort': 'medium'}
-> summary: "concise" is completely absent
Step 3 - handler hardcodes: {'effort': 'medium', 'summary': 'detailed'}
-> Expected summary: "concise", got: "detailed"
Component
SDK (litellm Python package)
LiteLLM Version
v1.81.10
What happened?
When using
litellm.aanthropic_messages()with a non-Anthropic model (e.g.openai/gpt-5.1) and passingthinking={"type": "enabled", "budget_tokens": 5000, "summary": "concise"}, thesummaryfield is silently dropped and replaced with a hardcoded"detailed".Expected: The user's
summary: "concise"should be preserved and sent to OpenAI's Responses API.Actual:
summaryis dropped bytranslate_anthropic_thinking_to_reasoning_effort()(returns plain string"medium"), then_route_openai_thinking_to_responses_api_if_needed()hardcodessummary: "detailed".Steps to Reproduce
Suggested Fix
The fix is small —
translate_anthropic_thinking_to_reasoning_effort()should return a dict instead of a string whensummaryis present in the thinking config:The handler already handles dicts correctly (lines 76-83 of
handler.py) — it only adds"summary": "detailed"when the key is absent.We'd be happy to open a PR for this if you'd accept one.
Workaround: We currently strip
thinkingfrom the kwargs and passreasoning_effortas a dict directly to bypass the translation.Relevant log output
Component
SDK (litellm Python package)
LiteLLM Version
v1.81.10