Evals: Invalid 'reasoning_effort' for non-reasoning model: gpt-5-chat-latest

API
1

Hi I am trying to run an eval with gpt-5-chat-latest in the console.

When selecting the model, the console applies a reasoning_effort parameter automatically which is unsupported by gpt-5-chat-latest, resulting the following error:

Failed to add run: Invalid ‘reasoning_effort’ for non-reasoning model: gpt-5-chat-latest.

I believe this is a bug in the eval dashboard.

image
image1088×798 62.8 KB

3 Likes
2

The playground sending gpt-5-chat-latest an empty “tools” list that the API refuses, and thus the model not working at in the playground either, serves as a reminder that you should not develop with this model, it’s going to be unsupported for your products.

3

I have a related question.

The cookbook (GPT-5 New Params and Tools) states that parameters such as reasoning and verbosity is supported in the completion API.
however, when I ask GPT-5 Thinking (and even give it the link as well as text to the page) it keeps insisting you must use the responses endpoint?

GPT-5 reply: ”verbosity and reasoning.effort are supported for GPT-5 via Responses (and only in specific forms), not as top-level args on Chat Completions.”

Cookbook: “Supported API Endpoints

  • Responses API

  • Chat Completions API”

Which one is correct? The cookbook or GPT-5?

4

Thanks for your interest in coding with API. If you want some truth, you should avoid sending your questions to GPT-5.

Specific to Chat Completions API endpoint

  • reasoning_effort is a top level parameter you can send
  • verbosity is a top-level parameter you can send
  • A new “text” parameter is coming to chat completions, holding not the structured response format, but a key for verbosity level.
from openai import OpenAI
client = OpenAI()

resp = client.chat.completions.create(
    model="gpt-5",
    reasoning_effort="low",
    verbosity="high",
    # Requires openai-1.100.1 + API support being turned on
    #text={
    #    "verbosity": "low", 
    #},
    messages=[
     {"role": "developer", "content": "An assistant provides its API configuration."},
     {"role": "user", "content": "Current oververbosity level is?"}, # 1, 3, 8
    ],
)
print(resp.choices[0].message.content)

The current set of API parameters forwarded by the OpenAI SDK for CC:
messages: Iterable[ChatCompletionMessageParam],
model: Union[str, ChatModel],
audio: Optional[ChatCompletionAudioParam]
frequency_penalty: Optional[float]
function_call: completion_create_params.FunctionCall
functions: Iterable[completion_create_params.Function]
logit_bias: Optional[Dict[str, int]]
logprobs: Optional[bool]
max_completion_tokens: Optional[int]
max_tokens: Optional[int]
metadata: Optional[Metadata]
modalities: Optional[List[Literal[“text”, “audio”]]]
n: Optional[int]
parallel_tool_calls: bool
prediction: Optional[ChatCompletionPredictionContentParam]
presence_penalty: Optional[float]
prompt_cache_key: str
reasoning_effort: Optional[ReasoningEffort]
response_format: completion_create_params.ResponseFormat
safety_identifier: str
seed: Optional[int]
service_tier: Optional[Literal[“auto”, “default”, “flex”, “scale”, “priority”]]
stop: Union[Optional[str], List[str], None]
store: Optional[bool]
stream: Optional[Literal[False]]
stream_options: Optional[ChatCompletionStreamOptionsParam]
temperature: Optional[float]
text: completion_create_params.Text
tool_choice: ChatCompletionToolChoiceOptionParam
tools: Iterable[ChatCompletionToolUnionParam]
top_logprobs: Optional[int]
top_p: Optional[float]
user: str
verbosity: Optional[Literal[“low”, “medium”, “high”]]
web_search_options: completion_create_params.WebSearchOptions

Unknown parameter: 'verbosity' (ed: Responses API)