Issue with reasoning tokens being used when using the "gpt-5.1-chat-latest" model

API Bugs
1

This is an issue discovered during testing using the recently available gpt-5.1-chat-latest model.
The documentation indicates that gpt-5.1-chat-latest does not support reasoning.

However, during testing, I confirmed that the reasoning token was used.

Used code

const openai = new OpenAI({ apiKey: process.env.AIAI_OPENAI_API_KEY });

(async () => {
  const result = await openai.responses.create({
    model: 'gpt-5.1-chat-latest',
    input: 'How much gold would it take to coat the Statue of Liberty in a 1mm layer?',
    reasoning: { effort: null },
  });

  console.log(result);
})();

Result

{
  usage: {
    input_tokens: 25,
    input_tokens_details: { cached_tokens: 0 },
    output_tokens: 179,
    output_tokens_details: { reasoning_tokens: 128 },
    total_tokens: 204
  }
}

I thought that when using the gpt-5.1-chat-latest model, reasoning_tokens should not be used. If I misunderstood something, please let me know.

Thank you.

4 Likes
Stop sequence doesn't work with gpt-5.1-chat-latest
2

Exactly
What is happening here?
This makes absolutely no sense

Why is the gpt-5.1-chat-latest a reasoning only model when the models page does not say it has reasoning?? (none of the chat-latest were ever reasoning)

Cant switch from something I had using completions with no reasoning on gpt-5-chat-latest into gpt-5.1-chat-latest

Now gpt-5.1-2025-11-13 supports reasoning_effort none, but gpt-5.1-chat-latest does not (only medium?? :confused: )

image
image1916Ă—827 82.2 KB

image
image1920Ă—952 86.5 KB

image
image1784Ă—862 86.4 KB

(also the same without using any libraries)

3

Were you guys able to discover what is happening? I’m also getting reasoning_tokens in the output tokens count.

5

nope, still confused about the reasoning of the chat-latest models since gpt-5.1-chat-latest
nobody seems to have an answer on this (not many seem to care)

6

The “chat” models are your copy of what runs on ChatGPT - running with that configuration. They don’t accept a reasoning effort parameter; logically you’d get the same internal reasoning (and pre-judgement) done on ChatGPT.

The actual reasoning tokens are obfuscated, <127 and you get billed 0.