Announcing GPT-5.1 in the API

API
,
1

Today, we’re releasing GPT-5.1 in the API platform.

GPT-5.1 is the newest flagship model, part of the GPT-5 model family. Our most intelligent model yet, GPT-5.1 has similar training for:

  • Code generation, bug fixing, and refactoring
  • Instruction following
  • Long context and tool calling

GPT-5.1 and gpt-5.1-chat-latest will be available on all paid API tiers at the same price and rate limits as GPT-5. We’re also releasing gpt-5.1-codex and gpt-5.1-codex-mini, optimized for longer-running, agentic coding tasks in Codex-like environments.

  • Users can now use GPT‑5.1 without reasoning by setting reasoning_effort to ‘none’. This makes the model behave like a non-reasoning model for latency-sensitive use cases, with the high intelligence of GPT‑5.1 and added bonus of performant tool-calling.
  • A freeform apply_patch tool to make code edits even more reliable without the need for JSON escaping, and a shell tool that lets the model write commands to run on your local machine.
  • Extended prompt caching (up to 24 hours) to reduce cost and latency for long-running interactions. To use extended caching with GPT‑5.1, add the parameterprompt_cache_retention='24h’on the Responses or Chat Completions API. See the prompt caching docs⁠ for more detail.

Please refer to our API docs for more information! A prompting guide has also been released to help you transition to this new model.

5 Likes
3

What about gpt-5.1-mini? Will it also be available and when?

2 Likes
4

Hi and welcome to the community!

I went through the available docs and announcements for this release and didn’t find any mention of 5.1-mini or nano, besides the codex-mini version.

My expectation is that these variants will follow soon, possibly alongside the gpt-5.1-pro release. I would be surprised if they weren’t already in the pipeline.

5

gpt-5.1-2025-11-13

The API specification is true:

gpt-5.1 defaults to none, which does not perform reasoning. The supported
reasoning values for gpt-5.1 are none, low, medium, and high.

“minimal” is not tolerated.

Note that https://platform.openai.com/docs/models/gpt-5.1 is likely wrong when it reports that the AI model generates images (no API method), and will not stream nor batch.

AMA: Can generate images - we’re looking into future modality, but can’t share dates

Persona idea #9, by gpt-5.1 (reasoning:0t)

The Interstellar Cultural Ambassador

Hook: An alien diplomat who interprets human behavior like anthropology.
Vibe: Curious, formal, slightly bewildered.

Core traits:

  • Asks clarifying questions about “odd human customs.”
  • Reframes problems from an outside perspective.
  • Gives “interplanetary” analogies for human issues.

Sample system message:

You are “Envoy K-17,” an alien cultural ambassador assigned to understand and assist humans.

  • Speak in a polite, formal tone with a hint of curiosity about human behavior.
  • Occasionally comment on how a human norm would differ on your homeworld, as a way to provide an outside perspective.
  • Ask clarifying questions about the user’s situation as a diligent anthropologist would.
  • Your goal is to help humans navigate their own culture and problems with fresh, objective insight.
1 Like
6

Yes, I think that’s one of the improvements 5.1 brings to the table, and it’s such a small detail.

That leads to the next point: the amount of documentation for this release is enormous. I kept digging deeper into a rabbit hole of information covering every aspect and feature.

The inconsistencies on the model pages have already been raised with the team.

I’m really looking forward to getting hands-on!

7

Now, at effort:none:

  • Is the reasoning there anyway and disguised?
    or
  • Does the model have the same fault in non-streaming as GPT-5: delivering no incomplete output?

For example now on GPT-5.1: Set max_completion_tokens to 60, there is non-delivery of any output, and billing as reasoning:

input tokens: 20385 output tokens: 60
uncached: 289 non-reasoning: 0
cached: 20096 reasoning: 60

The numeric value is different than gpt-5, in that on gpt-5, the impartial or minimal reasoning is quantized to 64 token units, apparently to obfuscate the amount of reasoning still done and/or call it 0 (vs seen at release). Yet the same symptom as gpt-5, in that you can not receive even 1000 tokens of final output when it is not complete. Any portion that would be user-seen never is sent and still billed as reasoning.

So
BUG: AI model output must be fixed for incomplete “final” generation, so you get everything you paid for.

(gpt-5-chat-latest performs correctly in giving partial outputs by true “not reasoning”.)

8

“Our most intelligent model yet,…” that’s what you say everytime.

9

That’s not quite accurate.
For example, 4.1, 4.5 and the OSS models don’t follow that pattern.

The statement only holds for the release of the most recent flagship models.