Merge pull request #26456 from BerriAI/litellm_hotfix_gpt-5.5-minimal-flag

litellm/llms/openai/chat/gpt_5_transformation.py

@@ -244,9 +244,11 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
                         ),
                         status_code=400,
                     )
-        elif effective_effort == "minimal":
-            # minimal is opt-out: unknown models pass through; only block when
-            # the model map explicitly sets supports_minimal_reasoning_effort=false.
+        elif effective_effort in ("minimal", "low"):
+            # minimal/low are opt-out: unknown models pass through; only block when
+            # the model map explicitly sets supports_{level}_reasoning_effort=false.
+            # Example: gpt-5.5-pro only accepts {medium, high, xhigh}, so it sets
+            # supports_low_reasoning_effort=false (and supports_minimal=false).
             if self._is_reasoning_effort_level_explicitly_disabled(
                 model, effective_effort
             ):

litellm/llms/vertex_ai/common_utils.py

@@ -656,6 +656,8 @@ def process_items(schema, depth=0):
             and ("items" not in schema or schema.get("items") == {})
         ):
             schema["items"] = {"type": "object"}
+        elif schema.get("type") == "array" and "items" not in schema:
+            schema["items"] = {"type": "object"}
         for key, value in schema.items():
             if isinstance(value, dict):
                 process_items(value, depth + 1)

litellm/model_prices_and_context_window_backup.json

@@ -19928,7 +19928,7 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": true,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false
     },
     "gpt-5.5-2026-04-23": {
         "cache_read_input_token_cost": 5e-07,
@@ -19976,7 +19976,7 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": true,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false
     },
     "gpt-5.5-pro": {
         "cache_read_input_token_cost": 3e-06,
@@ -20019,7 +20019,8 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": false,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false,
+        "supports_low_reasoning_effort": false
     },
     "gpt-5.5-pro-2026-04-23": {
         "cache_read_input_token_cost": 3e-06,
@@ -20062,7 +20063,8 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": false,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false,
+        "supports_low_reasoning_effort": false
     },
     "gpt-5.4": {
         "cache_read_input_token_cost": 2.5e-07,

litellm/types/utils.py

@@ -140,6 +140,7 @@ class ProviderSpecificModelInfo(TypedDict, total=False):
     supports_url_context: Optional[bool]
     supports_none_reasoning_effort: Optional[bool]
     supports_minimal_reasoning_effort: Optional[bool]
+    supports_low_reasoning_effort: Optional[bool]
     supports_xhigh_reasoning_effort: Optional[bool]
     supports_max_reasoning_effort: Optional[bool]
 

litellm/utils.py

@@ -5896,6 +5896,9 @@ def _get_model_info_helper(  # noqa: PLR0915
                 supports_minimal_reasoning_effort=_model_info.get(
                     "supports_minimal_reasoning_effort", None
                 ),
+                supports_low_reasoning_effort=_model_info.get(
+                    "supports_low_reasoning_effort", None
+                ),
                 supports_xhigh_reasoning_effort=_model_info.get(
                     "supports_xhigh_reasoning_effort", None
                 ),

model_prices_and_context_window.json

@@ -19942,7 +19942,7 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": true,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false
     },
     "gpt-5.5-2026-04-23": {
         "cache_read_input_token_cost": 5e-07,
@@ -19990,7 +19990,7 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": true,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false
     },
     "gpt-5.5-pro": {
         "cache_read_input_token_cost": 3e-06,
@@ -20033,7 +20033,8 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": false,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false,
+        "supports_low_reasoning_effort": false
     },
     "gpt-5.5-pro-2026-04-23": {
         "cache_read_input_token_cost": 3e-06,
@@ -20076,7 +20077,8 @@
         "supports_web_search": true,
         "supports_none_reasoning_effort": false,
         "supports_xhigh_reasoning_effort": true,
-        "supports_minimal_reasoning_effort": true
+        "supports_minimal_reasoning_effort": false,
+        "supports_low_reasoning_effort": false
     },
     "gpt-5.4": {
         "cache_read_input_token_cost": 2.5e-07,

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py

@@ -411,6 +411,46 @@ def test_generic_cost_per_token_gpt55_pro():
     )
 
 
+@pytest.mark.parametrize(
+    "model,expected_none,expected_xhigh,expected_minimal",
+    [
+        # Verified against OpenAI's live API on 2026-04-24:
+        #   gpt-5.5   -> supports: none, low, medium, high, xhigh
+        #   gpt-5.5-pro -> supports: medium, high, xhigh
+        # Neither supports "minimal"; gpt-5.5-pro additionally does not support "none".
+        # The JSON must reflect this so LiteLLM rejects unsupported values locally
+        # (or drops them with drop_params=True) instead of round-tripping to OpenAI
+        # for a 400.
+        ("gpt-5.5", True, True, False),
+        ("gpt-5.5-2026-04-23", True, True, False),
+        ("gpt-5.5-pro", False, True, False),
+        ("gpt-5.5-pro-2026-04-23", False, True, False),
+    ],
+)
+def test_gpt55_reasoning_effort_flags_match_live_openai_api(
+    model, expected_none, expected_xhigh, expected_minimal
+):
+    """Pin reasoning_effort capability flags to OpenAI's actual API contract.
+
+    Observed via `POST /v1/chat/completions` with reasoning_effort=minimal:
+    ``Unsupported value: 'reasoning_effort' does not support 'minimal' with
+    this model``. gpt-5.5-pro additionally rejects 'none' and 'low'.
+    """
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    m = litellm.model_cost[model]
+    assert (
+        m.get("supports_none_reasoning_effort") is expected_none
+    ), f"{model}: supports_none_reasoning_effort expected {expected_none}"
+    assert (
+        m.get("supports_xhigh_reasoning_effort") is expected_xhigh
+    ), f"{model}: supports_xhigh_reasoning_effort expected {expected_xhigh}"
+    assert (
+        m.get("supports_minimal_reasoning_effort") is expected_minimal
+    ), f"{model}: supports_minimal_reasoning_effort expected {expected_minimal}"
+
+
 @pytest.mark.parametrize(
     "base_model,dated_model",
     [

tests/test_litellm/llms/openai/test_gpt5_transformation.py

@@ -536,6 +536,81 @@ def test_gpt5_unknown_model_passes_through_minimal(config: OpenAIConfig):
     assert params["reasoning_effort"] == "minimal"
 
 
+def test_gpt5_5_pro_rejects_reasoning_effort_low(config: OpenAIConfig):
+    """gpt-5.5-pro only accepts {medium, high, xhigh} — 'low' must raise.
+
+    Verified against OpenAI's live API: /v1/chat/completions with
+    reasoning_effort='low' on gpt-5.5-pro returns HTTP 400.
+    """
+    with pytest.raises(litellm.utils.UnsupportedParamsError):
+        config.map_openai_params(
+            non_default_params={"reasoning_effort": "low"},
+            optional_params={},
+            model="gpt-5.5-pro",
+            drop_params=False,
+        )
+
+
+def test_gpt5_5_pro_dated_rejects_reasoning_effort_low(config: OpenAIConfig):
+    """Dated snapshot must inherit the base alias's low-rejection behavior."""
+    with pytest.raises(litellm.utils.UnsupportedParamsError):
+        config.map_openai_params(
+            non_default_params={"reasoning_effort": "low"},
+            optional_params={},
+            model="gpt-5.5-pro-2026-04-23",
+            drop_params=False,
+        )
+
+
+def test_gpt5_5_pro_drops_reasoning_effort_low_when_requested(config: OpenAIConfig):
+    """drop_params=True silently strips 'low' instead of round-tripping a 400."""
+    params = config.map_openai_params(
+        non_default_params={"reasoning_effort": "low"},
+        optional_params={},
+        model="gpt-5.5-pro",
+        drop_params=True,
+    )
+    assert "reasoning_effort" not in params
+
+
+def test_gpt5_5_chat_allows_reasoning_effort_low(config: OpenAIConfig):
+    """gpt-5.5 (chat) supports 'low'; flag absent → opt-out check passes."""
+    params = config.map_openai_params(
+        non_default_params={"reasoning_effort": "low"},
+        optional_params={},
+        model="gpt-5.5",
+        drop_params=False,
+    )
+    assert params["reasoning_effort"] == "low"
+
+
+def test_gpt5_unknown_model_passes_through_low(config: OpenAIConfig):
+    """Unknown gpt-5 models pass 'low' through (opt-out, not opt-in)."""
+    params = config.map_openai_params(
+        non_default_params={"reasoning_effort": "low"},
+        optional_params={},
+        model="gpt-5.4-turbo-preview",
+        drop_params=False,
+    )
+    assert params["reasoning_effort"] == "low"
+
+
+def test_gpt5_low_explicitly_disabled_check(gpt5_config: OpenAIGPT5Config):
+    """supports_low_reasoning_effort=false → disabled; missing/true → not disabled."""
+    assert gpt5_config._is_reasoning_effort_level_explicitly_disabled(
+        "gpt-5.5-pro", "low"
+    )
+    assert gpt5_config._is_reasoning_effort_level_explicitly_disabled(
+        "gpt-5.5-pro-2026-04-23", "low"
+    )
+    assert not gpt5_config._is_reasoning_effort_level_explicitly_disabled(
+        "gpt-5.5", "low"
+    )
+    assert not gpt5_config._is_reasoning_effort_level_explicitly_disabled(
+        "gpt-5.4", "low"
+    )
+
+
 def test_gpt5_normalizes_reasoning_effort_dict_with_summary(config: OpenAIConfig):
     """Dict with summary/generate_summary is normalized for chat completions."""
     params = config.map_openai_params(

tests/test_litellm/llms/vertex_ai/test_vertex_ai_common_utils.py

@@ -300,6 +300,18 @@ def test_process_items_basic():
     process_items(schema)
     assert schema["items"] == {"type": "object"}
 
+    # Test array missing items inside anyOf branch
+    schema = {
+        "type": "object",
+        "properties": {
+            "callbacks": {
+                "anyOf": [{"type": "array"}, {"type": "object"}],
+            }
+        },
+    }
+    process_items(schema)
+    assert schema["properties"]["callbacks"]["anyOf"][0]["items"] == {"type": "object"}
+
 
 def test_build_vertex_schema_array_branch_missing_items_in_anyof():
     """