fix: exclude gpt-5.2-chat from temperature passthrough

[giulio-leone]

Summary

drop_params: true does not drop temperature for gpt-5.2-chat / gpt-5.2-chat-latest models. These models only support temperature=1 (like base gpt-5), but were incorrectly classified as gpt-5.1/5.2 models that support arbitrary temperature.

Root Cause

is_model_gpt_5_1_model() uses model_name.startswith("gpt-5.2") which matches gpt-5.2-chat variants. These chat variants don't support arbitrary temperature (unlike gpt-5.2/gpt-5.2-codex).

Fix

Add not model_name.startswith("gpt-5.2-chat") to the is_gpt_5_2 classification check.

Test Changes

• Updated test_gpt5_1_model_detection to assert gpt-5.2-chat and gpt-5.2-chat-latest are NOT classified as gpt-5.1 models
• Added test_gpt5_2_chat_temperature_restricted regression test covering:
  • Raises UnsupportedParamsError for non-1 temperature
  • Allows temperature=1
  • drop_params=True silently drops non-1 temperature

All 31 tests pass.

Fixes #21911

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 28, 2026 4:17am

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR fixes an issue where gpt-5.2-chat and gpt-5.2-chat-latest models were incorrectly classified by is_model_gpt_5_1_model() as supporting arbitrary temperature values. The fix adds a not model_name.startswith("gpt-5.2-chat") check so these chat variants correctly fall through to the base gpt-5 temperature restriction (only temperature=1 allowed).

• Core fix in gpt_5_transformation.py is correct and minimal — adds a single exclusion condition to is_gpt_5_2 classification
• Test updates correctly verify gpt-5.2-chat variants are excluded from 5.1 model detection
• Regression test test_gpt5_2_chat_temperature_restricted covers error, passthrough, and drop_params scenarios
• Bug: The new test function accidentally absorbs the body of the former test_gpt5_2_pro_allows_reasoning_effort_xhigh test due to a bad merge/rebase — this standalone test is effectively deleted and its assertions are now orphaned inside an unrelated test function

Confidence Score: 3/5

• The core logic fix is correct and safe to merge, but the test file has a structural bug that should be fixed first.
• The production code change is a simple, correct one-line condition addition. However, the test file has an accidental merge issue where test_gpt5_2_pro_allows_reasoning_effort_xhigh was deleted and its body was absorbed into the new test function, which reduces test clarity and removes a named test case.
• tests/test_litellm/llms/openai/test_gpt5_transformation.py — the orphaned test code at lines 431-437 needs to be extracted back into its own test_gpt5_2_pro_allows_reasoning_effort_xhigh function.

Important Files Changed

Filename	Overview
litellm/llms/openai/chat/gpt_5_transformation.py	Correctly adds not model_name.startswith("gpt-5.2-chat") exclusion to is_model_gpt_5_1_model(), preventing gpt-5.2-chat variants from being classified as temperature-flexible models. Logic change is minimal and correct.
tests/test_litellm/llms/openai/test_gpt5_transformation.py	Good regression tests added, but the new test_gpt5_2_chat_temperature_restricted function accidentally absorbed the body of the old test_gpt5_2_pro_allows_reasoning_effort_xhigh test, which no longer exists as a separate function.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[map_openai_params called with temperature] --> B{is_model_gpt_5_1_model?}
    B -->|Yes| C{reasoning_effort == 'none' or None?}
    C -->|Yes| D[Allow any temperature]
    C -->|No| E{temperature == 1?}
    B -->|No| E
    E -->|Yes| F[Allow temperature=1]
    E -->|No| G{drop_params?}
    G -->|Yes| H[Silently drop temperature]
    G -->|No| I[Raise UnsupportedParamsError]

    subgraph "is_model_gpt_5_1_model (updated)"
        J[model_name] --> K{startswith gpt-5.1?}
        K -->|Yes| L[Return True]
        K -->|No| M{startswith gpt-5.2?}
        M -->|No| N[Return False]
        M -->|Yes| O{"'pro' in name?"}
        O -->|Yes| N
        O -->|No| P{"startswith gpt-5.2-chat? (NEW)"}
        P -->|Yes| N
        P -->|No| L
    end

Loading

_{Last reviewed commit: b57a908}

[greptile-apps]

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (1)

tests/test_litellm/llms/openai/test_gpt5_transformation.py
Orphaned test code merged into wrong function

The body of the former test_gpt5_2_pro_allows_reasoning_effort_xhigh test (lines 431-437) has been accidentally absorbed into test_gpt5_2_chat_temperature_restricted. In the base branch, this was a standalone test function; in the PR, the def line was replaced by the new test but the old body was left behind inside the new function.

The code still executes (it will pass), but it tests an unrelated concern (reasoning_effort='xhigh' for gpt-5.2-pro) inside a function that should only test gpt-5.2-chat temperature restrictions. The separate test_gpt5_2_pro_allows_reasoning_effort_xhigh test no longer exists.

These lines should be extracted back into their own test function:

        assert "temperature" not in params


def test_gpt5_2_pro_allows_reasoning_effort_xhigh(config: OpenAIConfig):
    params = config.map_openai_params(
        non_default_params={"reasoning_effort": "xhigh"},
        optional_params={},
        model="gpt-5.2-pro",
        drop_params=False,
    )
    assert params["reasoning_effort"] == "xhigh"