Merging this PR will not alter performance

✅ 16 untouched benchmarks

_{Comparing Sameerlite:litellm_gpt-5.4_mini (0dbed19) with main (cfeafbe)}

Greptile Summary

This PR adds Day-0 support for gpt-5.4-mini and gpt-5.4-nano by registering them in the model cost map and introducing a new minimal reasoning-effort level with opt-out semantics. It also fixes a pre-existing bug where the xhigh drop path only cleared non_default_params but not optional_params.

Key changes:

• New models: gpt-5.4-mini and gpt-5.4-nano added to both cost map files with supports_xhigh_reasoning_effort: true, supports_none_reasoning_effort: true, and supports_minimal_reasoning_effort: false.
• Bug fix (xhigh): The xhigh drop path now correctly pops from both non_default_params and optional_params, resolving the inconsistency flagged in the previous review.
• New minimal guard: An elif effective_effort == "minimal" branch uses the new _is_explicitly_disabled_factory opt-out helper — unknown models pass through, only explicit false entries are blocked.
• New helper: _is_explicitly_disabled_factory in litellm/utils.py mirrors _supports_factory but inverts the semantics (returns True only when the key is explicitly False).
• Concern: supports_minimal_reasoning_effort: true is added to ~25 pre-existing models. Under the opt-out semantics this is redundant (missing key = not disabled), but it also means these models will pass minimal through to OpenAI even if they do not actually support it. If any of those models (e.g. gpt-5-codex, gpt-5.1-codex) reject minimal on the live API, users will receive an upstream 400 rather than a clean LiteLLM error.

Confidence Score: 4/5

• Safe to merge after verifying that the ~25 existing models newly tagged with supports_minimal_reasoning_effort: true actually support minimal on the live OpenAI API.
• The core logic is correct and well-tested: the pre-existing xhigh optional_params drop bug has been fixed, the new minimal opt-out guard is sound, the _is_explicitly_disabled_factory helper is well-implemented, and all new tests are mock-only and comprehensive. The score is held at 4 (not 5) because supports_minimal_reasoning_effort: true is added to a large number of pre-existing models without per-model verification evidence — if any of those models reject minimal upstream, users will receive a raw OpenAI 400 instead of a clean LiteLLM error.
• Pay close attention to the supports_minimal_reasoning_effort: true additions in model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json for pre-existing models.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[map_openai_params called] --> B{effective_effort?}
    B -- xhigh --> C{_supports_reasoning_effort_level?\nopt-in: explicit True required}
    C -- supported --> D[pass xhigh through]
    C -- not supported --> E{drop_params?}
    E -- yes --> F[pop from non_default_params\nAND optional_params]
    E -- no --> G[raise UnsupportedParamsError]
    B -- minimal --> H{_is_reasoning_effort_level_explicitly_disabled?\nopt-out: explicit False required}
    H -- explicitly disabled\ni.e. supports_minimal=false in map --> I{drop_params?}
    I -- yes --> J[pop from non_default_params\nAND optional_params]
    I -- no --> K[raise UnsupportedParamsError]
    H -- not disabled\nmissing key OR true --> L[pass minimal through]
    B -- other / none / null --> M[continue normal flow]

_{Last reviewed commit: "Add test for reasoni..."}