fix(vertex_ai): don't send thinkingBudget: 0 for older Gemini models

Models with a minimum budget requirement (e.g. gemini-2.5-pro) reject
thinkingBudget: 0 at the API level. When budget_tokens=0, send
includeThoughts: False without a thinkingBudget field instead.

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -929,13 +929,17 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
                 # Thinking disabled
                 params["includeThoughts"] = False
         else:
-            # For older Gemini models, use thinkingBudget
-            if thinking_enabled and not VertexGeminiConfig._is_thinking_budget_zero(
-                thinking_budget
-            ):
-                params["includeThoughts"] = True
-            if thinking_budget is not None and isinstance(thinking_budget, int):
-                params["thinkingBudget"] = thinking_budget
+            # For older Gemini models, use thinkingBudget instead of thinkingLevel
+            if thinking_enabled:
+                if VertexGeminiConfig._is_thinking_budget_zero(thinking_budget):
+                    # thinkingBudget: 0 is rejected by models with a minimum budget (e.g. gemini-2.5-pro).
+                    params["includeThoughts"] = False
+                else:
+                    params["includeThoughts"] = True
+                    if thinking_budget is not None and isinstance(thinking_budget, int):
+                        params["thinkingBudget"] = thinking_budget
+            else:
+                params["includeThoughts"] = False
 
         return params
 

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -898,13 +898,16 @@ def test_vertex_ai_usage_metadata_accumulates_duplicate_modalities():
 
 def test_vertex_ai_map_thinking_param_with_budget_tokens_0():
     """
-    If budget_tokens is 0, do not set includeThoughts to True
+    budget_tokens=0 must produce includeThoughts: False with no thinkingBudget field.
+    Models with a minimum budget requirement (e.g. gemini-2.5-pro) reject thinkingBudget: 0.
     """
     from litellm.types.llms.anthropic import AnthropicThinkingParam
 
     v = VertexGeminiConfig()
     thinking_param: AnthropicThinkingParam = {"type": "enabled", "budget_tokens": 0}
-    assert "includeThoughts" not in v._map_thinking_param(thinking_param=thinking_param)
+    result = v._map_thinking_param(thinking_param=thinking_param)
+    assert result == {"includeThoughts": False}
+    assert "thinkingBudget" not in result
 
     thinking_param: AnthropicThinkingParam = {"type": "enabled", "budget_tokens": 100}
     assert v._map_thinking_param(thinking_param=thinking_param) == {
@@ -913,6 +916,19 @@ def test_vertex_ai_map_thinking_param_with_budget_tokens_0():
     }
 
 
+def test_vertex_ai_map_thinking_param_disabled():
+    """
+    type="adaptive" (non-enabled) must produce includeThoughts: False with no thinkingBudget field.
+    """
+    from litellm.types.llms.anthropic import AnthropicThinkingParam
+
+    v = VertexGeminiConfig()
+    thinking_param: AnthropicThinkingParam = {"type": "adaptive"}
+    result = v._map_thinking_param(thinking_param=thinking_param)
+    assert result == {"includeThoughts": False}
+    assert "thinkingBudget" not in result
+
+
 def test_vertex_ai_map_tools():
     v = VertexGeminiConfig()
     optional_params = {}