fix(anthropic_messages): gate sampling params on /v1/messages like /chat/completions

[mihidumh]

TLDR

Problem this solves:

• /v1/messages forwards temperature/top_p/top_k raw to models that removed them
• Provider 400s; router fallbacks mask them as silent model downgrades
• /chat/completions already drops the same params for the same models

How it solves it:

• Reuse the chat path's AnthropicModelInfo._apply_sampling_param gating in the /v1/messages param builder
• Drop under drop_params (matching /chat/completions), else clean client-side 400
• No-op for models that still support sampling params (temperature=1 stays allowed everywhere)

Relevant issues

Fixes #35053

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• 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
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

Screenshots / Proof of Fix

Real Vertex AI call (project-billed, no mocks), identical script before/after this change — litellm.anthropic_interface.acreate with model="vertex_ai/claude-sonnet-5", temperature=0.3, top_p=0.9, top_k=40, litellm.drop_params = True:

Before (base branch) — provider rejects the forwarded param:

litellm.llms.base_llm.chat.transformation.BaseLLMException: {"type":"error","error":
{"type":"invalid_request_error","message":"`temperature` is deprecated for this model."},
"request_id":"req_vrtx_011CdVrHWmLgPLyK75cwfNSC"}

After (this branch) — params dropped, request succeeds:

model: claude-sonnet-5 | stop_reason: end_turn
content: [{'type': 'text', 'text': 'Ok'}]

The same asymmetry was reproduced through a full proxy deployment (litellm 1.94.0 on the stock docker image) in #35053, including the silent-downgrade behavior when router_settings.fallbacks are configured — full 16-param × 2-endpoint × 2-generation sweep table there.

Unit tests: tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_request_optional_param_utils.py — drop for supports_sampling_params: false models (bare and vertex_ai/-prefixed ids), keep for supporting models, temperature=1 allowed, clean UnsupportedParamsError without drop_params. tests/test_litellm/llms/anthropic/ locally: 1184 passed vs 1179 on base, same 41 pre-existing env-dependent failures in both runs.

[greptile-apps]

Greptile Summary

This PR aligns Anthropic /v1/messages sampling-parameter handling with /chat/completions.

• Applies existing model-capability gating to temperature, top_p, and top_k.
• Drops unsupported parameters when configured and otherwise raises a client-side unsupported-parameter error.
• Adds coverage for unsupported, provider-prefixed, supporting, and temperature=1 cases.

Confidence Score: 5/5

The PR appears safe to merge, with the new messages-path behavior matching the established chat-path sampling rules.

The changed path delegates capability decisions and global or per-request drop behavior to the existing Anthropic helper, while preserving supported parameters and the allowed temperature value.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/utils.py	Reuses the existing Anthropic sampling-capability helper to filter or reject unsupported parameters without introducing a concrete regression.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_request_optional_param_utils.py	Adds focused unit coverage for dropping, preserving, and rejecting sampling parameters across relevant model forms.

_{Reviews (1): Last reviewed commit: "fix(anthropic_messages): gate sampling p..." | Re-trigger Greptile}

[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 mihidumh:fix/anthropic-messages-sampling-params (9653f58) with litellm_internal_staging (47a7e17)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (5290150) during the generation of this report, so 47a7e17 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mihidumh]

Closing — superseded by #37868, which merged this branch's commit (authorship preserved) into litellm_internal_staging on 2026-08-22. Thanks @tin-berri.