fix(anthropic): drop non-1 temperature when downgrading adaptive thinking

[axelray-dev]

Summary

When adaptive thinking is downgraded to legacy extended thinking for pre-4.6 Anthropic models, drop a non-1 temperature so the request is not rejected by Anthropic.

Changes

• In _translate_adaptive_effort_for_non_adaptive_model, after setting legacy thinking.type=enabled, remove temperature when it is present and not 1
• Leave temperature untouched for adaptive-capable models and when thinking is dropped entirely
• Add regression tests for Haiku 4.5 (drop/preserve) and Sonnet 4.6 (preserve temperature with adaptive)

Verification

• Focused tests in tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_effort.py

Backwards compatibility

No public API changes. Only strips an invalid temperature combination after adaptive-to-legacy translation.

Fixes #33203

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing axelray-dev:litellm_fix_anthropic_adaptive_temp_33203 (a74b353) with litellm_internal_staging (b200d66)}

[greptile-apps]

Greptile Summary

This PR fixes a rejection from Anthropic's API that occurs when a non-1 temperature is sent alongside legacy extended thinking (thinking.type=enabled) on pre-4.6 models. When the adaptive-to-legacy downgrade path produces a type=enabled thinking block, the fix now strips temperature if it is not 1.

• Two lines added in _translate_adaptive_effort_for_non_adaptive_model to drop non-1 temperature after setting legacy thinking.type=enabled; adaptive-capable models (4.6+) and the "drop thinking entirely" branch are deliberately unaffected.
• Three new unit tests cover the three distinct cases: temperature dropped for pre-4.6 legacy thinking, temperature=1 preserved, and temperature untouched for adaptive thinking on 4.6.

Confidence Score: 5/5

Safe to merge — the change is a two-line guard in a single downgrade path that can only remove a parameter Anthropic would reject anyway, and the new tests confirm all three temperature cases behave correctly.

The change is narrowly scoped to the adaptive-to-legacy translation path and cannot affect adaptive models or requests without thinking. The condition is logically correct for all temperature values (absent, 0, 1, 1.0), the pop default makes the absent-temperature case a no-op, and the three new tests directly exercise the regression being fixed. The only note is a minor readability improvement available on the condition guard.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/transformation.py	Adds two lines to drop temperature when it is not 1 during adaptive-to-legacy thinking downgrade; minor: the guard fires pop even when temperature is absent (harmless no-op due to default argument)
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_effort.py	Adds three focused unit tests (temperature=0 dropped, temperature=1 preserved for legacy thinking, temperature=0 preserved for adaptive thinking on 4.6); pure mock tests with no real network calls

_{Reviews (1): Last reviewed commit: "fix(anthropic): drop non-1 temperature w..." | Re-trigger Greptile}

[greptile-apps]

The condition optional_params.get("temperature") != 1 evaluates to True when temperature is absent (None != 1), so pop("temperature", None) is called on every legacy-thinking translation even when no temperature was set. The pop default prevents any error, but the intent — "only strip an explicitly-provided non-1 temperature" — is clearer with an explicit membership check.

Suggested change

	if capped_thinking.get("type") == "enabled" and optional_params.get("temperature") != 1:
	optional_params.pop("temperature", None)
	if (
	capped_thinking.get("type") == "enabled"
	and "temperature" in optional_params
	and optional_params["temperature"] != 1
	):
	optional_params.pop("temperature")

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[axelray-dev]

Closing this as superseded. The adaptive-to-legacy temperature drop already landed in #33244 (merged by krrish-berri-2), and #33203 is closed.