[BUG] Can't use GPT-5 model family - Unsupported parameter: `temperature` is not supported with this model.

[aperepel]

Bug Description

gpt-5-mini for model name, gpt-5-nano for small model. LLMConfig hardcodes the max_tokens parameter, which is rejected by gpt-5, as it replaced it with max_completion_tokens. The code is using a modern OpenAI client (not the OpenAI Generic compat one).

Steps to Reproduce

Provide a minimal code example that reproduces the issue:

llm_config = LLMConfig(
                    api_key=self._config['openai_api_key'],
                    base_url=self._config['openai_base_url'],
                    model=self._config['graphiti_model_name'],
                    small_model=self._config['graphiti_small_model_name'],
                    temperature=0.1
                    # omit max_tokens, it's hardcoded
                )

Expected Behavior

A clear and concise description of what you expected to happen.

Actual Behavior

Error on any attempt.

Environment

• Graphiti Version: 0.18.9
• OpenAI: 1.102.0
• LLM Provider & Model: [e.g. OpenAI gpt-4.1, Anthropic claude-4-sonnet, Google gemini-2.5-flash]

Installation Method

• pip install
• uv add
• Development installation (git clone)

Error Messages/Traceback

graphiti_core.llm_client.openai_base_client - ERROR - Error in generating LLM response: Error code: 400 - {'error': {'message': "Unsupported parameter: 'temperature' is not supported with this model.", 'type': 'invalid_request_error', 'param': 'temperature', 'code': None}}

Configuration

# Relevant configuration or initialization code

Additional Context

• Does this happen consistently or intermittently?
• Which component are you using? (core library, REST server, MCP server)
• Any recent changes to your environment?
• Related issues or similar problems you've encountered?

Possible Solution

If you have ideas about what might be causing the issue or how to fix it, please share them here.

[prasmussen15]

If you are using gpt-5 models (the new defaults) you can't pass in a temperature other than 1 as the gpt-5 models are thinking models and so only support temperature=1. The max_completion_tokens is handled correctly.

[aperepel]

Sorry for the confusion, the github issue editor somehow kept an old version of the text.

The issue is not the max_tokens, this has been addressed in the latest graphiti versions from what I can tell. The current issue is, the temperature param is always passed in and, subsequently, rejected by the API.

An interesting observation: calling via OpenRouter APIs works, it must be transforming the API call data and stripping incompatible args. Calling OpenAI API directly exhibits the 400 error complaining about the temperature arg.

Can you please reopen the issue?

Here's the workaround I had to put in to control (remove) the arg:

class GPT5CompatibleOpenAIClient(OpenAIClient):
    """
    OpenAI client that excludes temperature parameter for GPT-5 compatibility.
    
    GPT-5 doesn't support the temperature parameter, so this client overrides
    the completion methods to exclude it from API calls.
    """
    
    async def _create_structured_completion(
        self,
        model: str,
        messages,
        temperature: float | None,
        max_tokens: int,
        response_model,
    ):
        """Create a structured completion without temperature parameter."""
        response = await self.client.responses.parse(
            model=model,
            input=messages,  # type: ignore
            max_output_tokens=max_tokens,
            text_format=response_model,  # type: ignore
            # Deliberately exclude temperature parameter
        )
        return response

    async def _create_completion(
        self,
        model: str,
        messages,
        temperature: float | None,
        max_tokens: int,
        response_model = None,
    ):
        """Create a regular completion without temperature parameter."""
        return await self.client.chat.completions.create(
            model=model,
            messages=messages,
            max_tokens=max_tokens,
            response_format={'type': 'json_object'},
            # Deliberately exclude temperature parameter
        )

[prasmussen15]

In your example you are passing in a value for temperature but are using the gpt-5 family of models (the defaults model types). The gpt-5 models only support temperature=1 so your options are to either pass in temeprature=1 or not pass in a value for temperature.

Stripping the temperature when it is explicitly passed in is an anti-pattern as it is a silent no-op on the temperature parameter. Here the explicit error is letting you know that temperature=.1 is not a supported value for these models.

[weikhang95]

@prasmussen15 I still having issue with the max_completion_tokens. I am using gpt-5-mini.

2025-08-29 10:49:01,240 - httpx - INFO - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 400 Bad Request"
2025-08-29 10:49:01,240 - graphiti_core.llm_client.openai_base_client - ERROR - Error in generating LLM response: Error code: 400 - {'error': {'message': "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.", 'type': 'invalid_request_error', 'param': 'max_tokens', 'code': 'unsupported_parameter'}}

[prasmussen15]

@weikhang95 what version of graphiti and which llm_client are you using?

[freedomwarrior]

Hello.
Got the same issue

"Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead."

I'm using latest docker image

zepai/knowledge-graph-mcp