fix(responses-api): apply GPT-5 temperature validation

[Chesars]

Relevant issues

Fixes #16090

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

The Responses API map_openai_params passed all params through without model-specific validation. GPT-5 models (except gpt-5-chat) only accept temperature=1 unless reasoning.effort="none" on models that support it (5.1, 5.2, 5.4). The chat completions path already validates this, but the Responses API path did not.

Tests added

• test_responses_gpt5_drop_temperature — drop_params drops temperature!=1
• test_responses_gpt5_reject_temperature — raises error without drop_params
• test_responses_gpt5_allow_temperature_1 — temperature=1 always allowed
• test_responses_gpt5_mini_drop_temperature — gpt-5-mini also drops
• test_responses_gpt5_chat_allow_temperature — gpt-5-chat not restricted

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 22, 2026 9:39pm

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing Chesars:fix/responses-api-gpt5-temperature-drop-params (bfee7f0) with main (c89496f)}

[greptile-apps]

Greptile Summary

This PR adds GPT-5 temperature validation to the Responses API map_openai_params path, fixing a gap where the chat completions path already enforced that GPT-5 models (except gpt-5-chat) only accept temperature=1 — unless reasoning.effort='none' on models like gpt-5.1/5.4 that support it — but the Responses API path passed all parameters through without any validation.

Changes:

• litellm/llms/openai/responses/transformation.py: map_openai_params now reuses OpenAIGPT5Config.is_model_gpt_5_model and _supports_reasoning_effort_level to apply the same temperature guard that exists in the chat completions path; raises UnsupportedParamsError or drops the param according to drop_params.
• tests/test_litellm/llms/openai/test_gpt5_transformation.py: 7 new pure unit tests covering drop, reject, allow (temperature=1), gpt-5-mini, gpt-5-chat bypass, gpt-5.1 no-effort, and gpt-5.1/5.4 with explicit effort values.

Notes:

• The fix correctly mirrors the existing chat completions logic, including the effort is None → allows temperature semantic for supports-none models.
• is_model_gpt_5_model relies on a hardcoded substring check rather than the model map (model_prices_and_context_window.json), which is an existing pattern but now applied to the Responses API path as well. A future follow-up using _supports_factory for temperature restriction detection would be more future-proof and consistent with the project's custom rule.
• This change is technically a behavior shift: callers previously received an API-level HTTP 400; they now receive litellm.UnsupportedParamsError before the network call is made. Existing code catching openai.BadRequestError should be updated to also handle litellm.UnsupportedParamsError.

Confidence Score: 4/5

• Safe to merge — the fix is correct, well-tested, and consistent with the existing chat completions path; only minor style/future-proofing concerns remain.
• The core logic faithfully mirrors the established chat completions path, unit test coverage is thorough (no network calls, multiple model variants), and the error path is validated. The score is not 5 because the hardcoded model-detection helper (is_model_gpt_5_model) deviates from the project's preferred model-map-driven pattern, and the exception-type change (litellm-layer error vs. API-layer error) could silently break existing callers' exception handling.
• No files require special attention; litellm/llms/openai/responses/transformation.py is the only logic change and it is straightforward.

Important Files Changed

Filename	Overview
litellm/llms/openai/responses/transformation.py	Adds GPT-5 temperature validation to map_openai_params, mirroring the chat completions path. Logic is correct and consistent; minor concern around hardcoded model detection (reusing existing helper) and a behavior-level change in exception type (UnsupportedParamsError raised earlier instead of API-level 400).
tests/test_litellm/llms/openai/test_gpt5_transformation.py	Adds 7 new unit tests covering drop, reject, and allow paths for several model variants. All tests are pure in-memory (no network calls), satisfying the test-folder constraint. Missing a reject-path test for supports-none models (e.g. gpt-5.1 + effort='high' + drop_params=False), but this was noted in previous review threads.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[map_openai_params called] --> B{is_model_gpt_5_model?}
    B -- No --> Z[return params as-is]
    B -- Yes --> C{temperature in params\nAND temperature != 1?}
    C -- No --> Z
    C -- Yes --> D[extract reasoning.effort]
    D --> E{_supports_reasoning_effort_level\nlevel='none'?}
    E -- Yes, supports_none=True --> F{effort == 'none'\nOR effort is None?}
    F -- Yes --> Z[return params with temperature]
    F -- No --> G{drop_params\nOR litellm.drop_params?}
    E -- No, supports_none=False --> G
    G -- Yes --> H[pop temperature from params]
    H --> Z
    G -- No --> I[raise UnsupportedParamsError 400]

Loading

_{Reviews (3): Last reviewed commit: "fix: improve error message for supports-..." | Re-trigger Greptile}