fix(anthropic/passthrough): drop incompatible temperature when downgrading adaptive thinking for pre-4.6 models

[devin-ai-integration]

Relevant issues

Fixes #33203

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem

Screenshots / Proof of Fix

Live proxy against the real Anthropic API, same request both times (claude-haiku-4-5, thinking: {"type": "adaptive"}, temperature: 0, max_tokens: 8192, run with LITELLM_LOCAL_MODEL_COST_MAP=True so Haiku 4.5 registers supports_reasoning)

Before the fix (parent 93b5ca9612), the adaptive interface is downgraded to legacy enabled thinking but temperature stays pinned, so Anthropic rejects it:

$ curl -sw "\nHTTP %{http_code}\n" -X POST http://localhost:4000/v1/messages \
  -H "Authorization: Bearer sk-1234" -H "content-type: application/json" \
  -d '{"model":"claude-haiku-4-5","max_tokens":8192,"temperature":0,"thinking":{"type":"adaptive"},"messages":[{"role":"user","content":"Reply with the single word: ok"}]}'

{"error":{"message":"...invalid_request_error...`temperature` may only be set to 1 when thinking is enabled...","code":"400"}}
HTTP 400

After the fix (c4413e91f3), the same request succeeds and thinking is preserved:

{"model":"claude-haiku-4-5","type":"message","role":"assistant",
 "content":[{"type":"thinking","thinking":"The user is asking me to reply with just the single word \"ok\"...","signature":"..."},
            {"type":"text","text":"ok"}], ...}
HTTP 200

The proxy --detailed_debug log confirms the outgoing body to api.anthropic.com carries the downgraded thinking and no temperature:

POST Request Sent from LiteLLM:
https://api.anthropic.com/v1/messages
-d '{'messages': [...], 'max_tokens': 8192, 'model': 'claude-haiku-4-5', 'stream': False, 'thinking': {'type': 'enabled', 'budget_tokens': 2048}}'

The temperature-dropping logic is unchanged since c4413e91f3, so this proof still holds at the current head

Type

🐛 Bug Fix

Changes

On the native Anthropic /v1/messages passthrough, clients like Claude Code send thinking: {"type": "adaptive"} unconditionally. When such a request lands on a pre-4.6 model (e.g. claude-haiku-4-5), _translate_adaptive_effort_for_non_adaptive_model downgrades the adaptive interface to legacy thinking={type: enabled, budget_tokens}, but it left temperature untouched. Anthropic then rejects the request with 400 invalid_request_error: "temperature may only be set to 1 when thinking is enabled". This bites Claude Code specifically because its permission-safety classifier pins temperature: 0 on Haiku 4.5 for determinism, so the first classifier action 400s

The fix adds _drop_incompatible_temperature_for_thinking, run after all thinking translations in transform_anthropic_messages_request. For non-adaptive models, when the reshaped request still enables extended thinking (either legacy thinking.type == "enabled" or a kept output_config.effort, as on Opus 4.5) and the caller pinned temperature to a value other than 1, it drops the temperature so Anthropic applies its default of 1. Preserving the thinking the request asked for wins over an unhonorable sampling value, since Anthropic forces temperature=1 under thinking regardless. Adaptive models (4.6+) own this natively and are left untouched, so the main Claude Code loop is unaffected

Pseudocode:

def _drop_incompatible_temperature_for_thinking(model, optional_params, provider):
    if is_adaptive_thinking_model(model, provider):
        return
    temp = optional_params.get("temperature")
    if temp is None or temp == 1:
        return
    thinking_enabled = optional_params.get("thinking", {}).get("type") == "enabled"
    effort_enabled = optional_params.get("output_config", {}).get("effort") is not None
    if thinking_enabled or effort_enabled:
        optional_params.pop("temperature", None)

Scope note on the sibling defect: the issue also flagged that _translate_reasoning_effort_to_anthropic can emit budget_tokens >= max_tokens. An earlier revision of this PR capped the budget below max_tokens, but that regressed the live translation grid (tests/llm_translation/reasoning_effort_grid): budget-mode models advertise no supports_xhigh/supports_max_reasoning_effort capability, and the grid relies on the resulting budget_tokens >= max_tokens to make Anthropic reject xhigh/max with a 400. Silently capping turned those 400s into 200s, honoring an effort tier the model does not support. Fixing that defect correctly needs a capability gate for the non-adaptive legacy path (reject unsupported tiers explicitly, then cap the supported ones), which is a larger, separate change. This PR is scoped to the temperature bug in #33203 and leaves the budget behavior at the current baseline

All changes are in litellm/llms/anthropic/experimental_pass_through/messages/transformation.py plus regression tests in tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_effort.py

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/58efcf465eb24622b768fe3905c9a781

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes two related defects in the Anthropic passthrough path for pre-4.6 models. When thinking: {type: "adaptive"} is downgraded to legacy thinking: {type: "enabled"}, a co-present temperature != 1 would cause Anthropic to return a 400; a new _drop_incompatible_temperature_for_thinking step silently removes the conflicting temperature after all thinking translations complete. A sibling fix caps budget_tokens below max_tokens in _translate_reasoning_effort_to_anthropic, matching the cap already applied in the adaptive-downgrade path.

• Adds _drop_incompatible_temperature_for_thinking, run last in transform_anthropic_messages_request, that drops temperature when legacy extended thinking or native output_config.effort is active on non-adaptive models; adaptive 4.6+ models are untouched.
• Updates _translate_reasoning_effort_to_anthropic to apply _cap_thinking_budget_to_max_tokens before setting thinking, preventing budget_tokens >= max_tokens rejections.
• Existing test test_reasoning_effort_on_non_adaptive_model_uses_thinking_budget is correctly updated: the old assertion (budget_tokens >= 1024) was asserting the buggy uncapped behaviour; nine new regression tests cover the fixed behaviour and all relevant edge cases.

Confidence Score: 4/5

Safe to merge. The changes are well-scoped to the passthrough transformation layer, all new code paths are covered by unit tests, and no production-path logic is broken.

The temperature-drop and budget-cap logic are straightforward and well-tested. The DROP_UNSUPPORTED_ADAPTIVE_EFFORT_WARNING message reused in _translate_reasoning_effort_to_anthropic (line 231) says "adaptive" when the actual cause is a reasoning_effort-derived budget failing the max_tokens cap, which could mislead operators in logs. No functional bugs were found beyond this observability gap.

transformation.py around the _translate_reasoning_effort_to_anthropic warning path (line 231) — the reused adaptive-effort warning message is inaccurate for the reasoning_effort code path.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/transformation.py	Adds _drop_incompatible_temperature_for_thinking to silently drop non-1 temperature when extended thinking is active on non-adaptive models, and adds budget capping to the _translate_reasoning_effort_to_anthropic path. Logic is sound; minor warning-message reuse is slightly misleading in logs.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_effort.py	Adds nine new regression tests covering temperature-drop scenarios, budget-capping, and per-model edge cases. All assertions are correct and use mock-only paths.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_reasoning_effort_translation.py	Existing test test_reasoning_effort_on_non_adaptive_model_uses_thinking_budget legitimately updated: old max_tokens=1024 asserted the buggy uncapped behavior; new max_tokens=8192 asserts the corrected capped behavior and is accompanied by a dedicated boundary-case test.

_{Reviews (1): Last reviewed commit: "fix(anthropic/passthrough): drop tempera..." | Re-trigger Greptile}

[greptile-apps]

Misleading warning message reused from adaptive path

DROP_UNSUPPORTED_ADAPTIVE_EFFORT_WARNING says "Dropping adaptive thinking/output_config.effort…" but here the thinking being dropped was derived from the reasoning_effort OpenAI alias, not from an adaptive interface. An operator seeing this warning in logs while using reasoning_effort will have a hard time understanding the root cause — max_tokens too small for the translated budget is the real reason, not an adaptive-thinking rejection.

[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 litellm_fix_33203_passthrough_thinking_temperature (fbb2c1c) with litellm_internal_staging (8b32320)}