Fix: Apply additional_drop_params filtering across all LLM providers

- Added centralized _apply_additional_drop_params helper method in BaseLLM
- Applied filtering in LiteLLM fallback path (_prepare_completion_params in llm.py)
- Applied filtering in Azure native provider (azure/completion.py)
- Applied filtering in OpenAI native provider (openai/completion.py)
- Applied filtering in Anthropic native provider (anthropic/completion.py)
- Added comprehensive tests covering both direct LLM usage and Agent usage
- Tests verify filtering works for single and multiple parameters
- Tests verify backwards compatibility with misspelled drop_additionnal_params

Fixes #3814

Co-Authored-By: João <joao@crewai.com>

lib/crewai/src/crewai/llm.py

@@ -498,7 +498,11 @@ class LLM(BaseLLM):
         }
 
         # Remove None values from params
-        return {k: v for k, v in params.items() if v is not None}
+        params = {k: v for k, v in params.items() if v is not None}
+
+        params = self._apply_additional_drop_params(params)
+
+        return params
 
     def _handle_streaming_response(
         self,

lib/crewai/src/crewai/llms/base_llm.py

@@ -223,6 +223,49 @@ class BaseLLM(ABC):
 
         return content
 
+    def _apply_additional_drop_params(self, params: dict[str, Any]) -> dict[str, Any]:
+        """Apply additional_drop_params filtering to remove unwanted parameters.
+
+        This method provides consistent parameter filtering across all LLM providers.
+        It should be called after building the final params dict and before making
+        the provider API call.
+
+        Args:
+            params: The parameters dictionary to filter
+
+        Returns:
+            Filtered parameters dictionary with specified params removed
+
+        Example:
+            >>> llm = LLM(model="o1-mini", additional_drop_params=["stop"])
+            >>> params = {"model": "o1-mini", "messages": [...], "stop": ["\\n"]}
+            >>> filtered = llm._apply_additional_drop_params(params)
+            >>> "stop" in filtered
+            False
+        """
+        drop_params = (
+            self.additional_params.get("additional_drop_params")
+            or self.additional_params.get("drop_additionnal_params")
+            or []
+        )
+
+        if not drop_params:
+            return params
+
+        filtered_params = params.copy()
+
+        for param_name in drop_params:
+            if param_name in filtered_params:
+                logging.debug(
+                    f"Dropping parameter '{param_name}' as specified in additional_drop_params"
+                )
+                filtered_params.pop(param_name)
+
+        filtered_params.pop("additional_drop_params", None)
+        filtered_params.pop("drop_additionnal_params", None)
+
+        return filtered_params
+
     def get_context_window_size(self) -> int:
         """Get the context window size for the LLM.
 

lib/crewai/src/crewai/llms/providers/anthropic/completion.py

@@ -201,6 +201,8 @@ class AnthropicCompletion(BaseLLM):
         if tools and self.supports_tools:
             params["tools"] = self._convert_tools_for_interference(tools)
 
+        params = self._apply_additional_drop_params(params)
+
         return params
 
     def _convert_tools_for_interference(self, tools: list[dict]) -> list[dict]:

lib/crewai/src/crewai/llms/providers/azure/completion.py

@@ -273,6 +273,8 @@ class AzureCompletion(BaseLLM):
             params["tools"] = self._convert_tools_for_interference(tools)
             params["tool_choice"] = "auto"
 
+        params = self._apply_additional_drop_params(params)
+
         return params
 
     def _convert_tools_for_interference(self, tools: list[dict]) -> list[dict]:

lib/crewai/src/crewai/llms/providers/openai/completion.py

@@ -249,7 +249,11 @@ class OpenAICompletion(BaseLLM):
             "timeout",
         }
 
-        return {k: v for k, v in params.items() if k not in crewai_specific_params}
+        params = {k: v for k, v in params.items() if k not in crewai_specific_params}
+
+        params = self._apply_additional_drop_params(params)
+
+        return params
 
     def _convert_tools_for_interference(self, tools: list[dict]) -> list[dict]:
         """Convert CrewAI tool format to OpenAI function calling format."""

lib/crewai/tests/test_llm.py

@@ -725,3 +725,108 @@ def test_native_provider_falls_back_to_litellm_when_not_in_supported_list():
         # Should fall back to LiteLLM
         assert llm.is_litellm is True
         assert llm.model == "groq/llama-3.1-70b-versatile"
+
+
+def test_additional_drop_params_filters_parameters_in_litellm():
+    """Test that additional_drop_params correctly filters out specified parameters in LiteLLM path."""
+    with patch("crewai.llm.LITELLM_AVAILABLE", True), patch("crewai.llm.litellm"):
+        llm = LLM(
+            model="o1-mini",
+            stop=["stop_sequence"],
+            additional_drop_params=["stop"],
+            is_litellm=True,
+        )
+
+        messages = [{"role": "user", "content": "Hello"}]
+        params = llm._prepare_completion_params(messages)
+
+        assert "stop" not in params
+        assert "additional_drop_params" not in params
+        assert params["model"] == "o1-mini"
+
+
+def test_additional_drop_params_with_agent():
+    """Test that additional_drop_params works when LLM is used with an Agent."""
+    from unittest.mock import patch, MagicMock
+    from crewai import Agent, Task, Crew
+
+    with patch("crewai.llm.LITELLM_AVAILABLE", True), patch("crewai.llm.litellm") as mock_litellm:
+        llm = LLM(
+            model="o1-mini",
+            stop=["stop_sequence"],
+            additional_drop_params=["stop"],
+            is_litellm=True,
+        )
+
+        agent = Agent(
+            role="Test Agent",
+            goal="Test the LLM response format functionality.",
+            backstory="An AI developer testing LLM integrations.",
+            llm=llm,
+        )
+
+        task = Task(
+            description="Say hello",
+            expected_output="A greeting",
+            agent=agent,
+        )
+
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "Hello!"
+        mock_response.choices[0].message.tool_calls = None
+        mock_response.usage = MagicMock()
+        mock_response.usage.prompt_tokens = 10
+        mock_response.usage.completion_tokens = 5
+        mock_response.usage.total_tokens = 15
+        mock_litellm.completion.return_value = mock_response
+
+        crew = Crew(agents=[agent], tasks=[task], verbose=False)
+        crew.kickoff()
+
+        # Verify that litellm.completion was called
+        assert mock_litellm.completion.called
+
+        # Get the kwargs passed to litellm.completion
+        call_kwargs = mock_litellm.completion.call_args[1]
+
+        assert "stop" not in call_kwargs
+        assert "additional_drop_params" not in call_kwargs
+
+
+def test_additional_drop_params_supports_misspelled_variant():
+    """Test that drop_additionnal_params (misspelled) is also supported for backwards compatibility."""
+    with patch("crewai.llm.LITELLM_AVAILABLE", True), patch("crewai.llm.litellm"):
+        llm = LLM(
+            model="o1-mini",
+            stop=["stop_sequence"],
+            drop_additionnal_params=["stop"],
+            is_litellm=True,
+        )
+
+        messages = [{"role": "user", "content": "Hello"}]
+        params = llm._prepare_completion_params(messages)
+
+        assert "stop" not in params
+        assert "drop_additionnal_params" not in params
+        assert params["model"] == "o1-mini"
+
+
+def test_additional_drop_params_filters_multiple_parameters():
+    """Test that additional_drop_params can filter multiple parameters."""
+    with patch("crewai.llm.LITELLM_AVAILABLE", True), patch("crewai.llm.litellm"):
+        llm = LLM(
+            model="o1-mini",
+            stop=["stop_sequence"],
+            temperature=0.7,
+            additional_drop_params=["stop", "temperature"],
+            is_litellm=True,
+        )
+
+        messages = [{"role": "user", "content": "Hello"}]
+        params = llm._prepare_completion_params(messages)
+
+        assert "stop" not in params
+        assert "temperature" not in params
+        assert "additional_drop_params" not in params
+        assert params["model"] == "o1-mini"