diff --git a/benchmark/llm_client.py b/benchmark/llm_client.py
index 46d64e5..b393c00 100644
--- a/benchmark/llm_client.py
+++ b/benchmark/llm_client.py
@@ -35,6 +35,8 @@ def __init__(
         self.top_p = top_p
         self.effort = effort
         self.reasoning = reasoning
+        # Responses API only when an effort is set for OpenAI/AzureOpenAI endpoints.
+        self.use_responses_api = bool(effort) and self.provider in ("openai", "azureopenai")
         self.llm = None
 
         self._initialize_llm()
@@ -65,11 +67,18 @@ def _initialize_llm(self):
             elif self.provider == "openai":
                 from langchain_openai import ChatOpenAI
 
+                model_kwargs = {}
+                if self.top_p is not None:
+                    model_kwargs["top_p"] = self.top_p
+
                 self.llm = ChatOpenAI(
                     model=self.model,
                     openai_api_key=self.api_key,
                     temperature=self.temperature,
                     max_tokens=self.max_tokens,
+                    model_kwargs=model_kwargs,
+                    use_responses_api=self.use_responses_api,
+                    reasoning_effort=self.effort,
                 )
             elif self.provider == "google":
                 from langchain_google_genai import ChatGoogleGenerativeAI
@@ -97,8 +106,6 @@ def _initialize_llm(self):
                 model_kwargs = {}
                 if self.top_p is not None:
                     model_kwargs["top_p"] = self.top_p
-                if self.effort is not None:
-                    model_kwargs["reasoning_effort"] = self.effort
 
                 self.llm = AzureChatOpenAI(
                     azure_endpoint=self.custom_api_endpoint,
@@ -107,7 +114,9 @@ def _initialize_llm(self):
                     azure_deployment=self.model,
                     temperature=self.temperature,
                     max_completion_tokens=self.max_tokens,
-                    model_kwargs=model_kwargs # In GPT-5.X this is a first class parameter, but passing this way is also allowed.
+                    model_kwargs=model_kwargs,
+                    use_responses_api=self.use_responses_api, # Required for some reasoning model configurations
+                    reasoning_effort=self.effort,
                 )
             elif self.provider == "vllm" or self.provider == "openrouter":
                 from langchain_openai import ChatOpenAI
@@ -200,6 +209,10 @@ def _convert_mcp_tools_to_langchain(
                     "parameters": cleaned_schema,
                 },
             }
+            # Responses API defaults to strict mode, which forces the model to fill every
+            # optional param with hallucinated values. Explicit strict=False avoids this.
+            if self.provider in ("openai", "azureopenai") and self.use_responses_api:
+                tool_def["function"]["strict"] = False
             langchain_tools.append(tool_def)
 
         return langchain_tools
@@ -314,7 +327,8 @@ async def invoke_with_tools(
         # Convert MCP tools to LangChain format
         langchain_tools = self._convert_mcp_tools_to_langchain(tools)
 
-        # Bind tools to LLM
+        # Bind tools to LLM (the strict=False flag for OpenAI providers is
+        # set on each tool dict in _convert_mcp_tools_to_langchain).
         llm_with_tools = self.llm.bind_tools(langchain_tools)
         llm_with_retry = llm_with_tools.with_retry(
             retry_if_exception_type=(