diff --git a/rdagent/oai/backend/deprec.py b/rdagent/oai/backend/deprec.py
index e5601d49..a6caeed8 100644
--- a/rdagent/oai/backend/deprec.py
+++ b/rdagent/oai/backend/deprec.py
@@ -194,7 +194,7 @@ class ChatSession:
         history_message = SessionChatHistoryCache().message_get(self.conversation_id)
         messages = history_message
         if not messages:
-            messages.append({"role": "system", "content": self.system_prompt})
+            messages.append({"role": LLM_SETTINGS.system_prompt_role, "content": self.system_prompt})
         messages.append(
             {
                 "role": "user",
@@ -477,7 +477,7 @@ class DeprecBackend(APIBackend):
         system_prompt = LLM_SETTINGS.default_system_prompt if system_prompt is None else system_prompt
         messages = [
             {
-                "role": "system",
+                "role": LLM_SETTINGS.system_prompt_role,
                 "content": system_prompt,
             },
         ]
@@ -799,7 +799,8 @@ class DeprecBackend(APIBackend):
                 if add_json_in_prompt:
                     for message in messages[::-1]:
                         message["content"] = message["content"] + "\nPlease respond in json format."
-                        if message["role"] == "system":
+                        if message["role"] == LLM_SETTINGS.system_prompt_role:
+                            # NOTE: assumption: systemprompt is always the first message
                             break
                 call_kwargs["response_format"] = {"type": "json_object"}
             response = self.chat_client.chat.completions.create(**call_kwargs)
diff --git a/rdagent/oai/llm_conf.py b/rdagent/oai/llm_conf.py
index 90341db6..34c564e9 100644
--- a/rdagent/oai/llm_conf.py
+++ b/rdagent/oai/llm_conf.py
@@ -11,6 +11,9 @@ class LLMSettings(ExtendedBaseSettings):
     # backend
     backend: str = "rdagent.oai.backend.DeprecBackend"
 
+    # TODO: most of the settings are only used on deprec.DeprecBackend.
+    # So they should move the settings to that folder.
+
     log_llm_chat_content: bool = True
 
     use_azure: bool = Field(default=False, deprecated=True)
@@ -56,6 +59,9 @@ class LLMSettings(ExtendedBaseSettings):
         100000  # 100000 is the maximum limit of gpt4, which might increase in the future version of gpt
     )
     default_system_prompt: str = "You are an AI assistant who helps to answer user's questions."
+    system_prompt_role: str = "system"
+    """Some models (like o1) do not support the 'system' role.
+    Therefore, we make the system_prompt_role customizable to ensure successful calls."""
 
     # Embedding configs
     embedding_openai_api_key: str = ""