OpenAI releases GPT-4.1 in the API

OpenAI released GPT-4.1 as a family of three models — GPT-4.1, GPT-4.1 mini and GPT-4.1 nano — available only through the API and developer playground, not in ChatGPT. All three carried a context window OpenAI put at one million tokens, and the family was pitched primarily on coding and instruction-following gains over GPT-4o, including a reported improvement of more than 20 percentage points on SWE-bench Verified, a benchmark measuring the ability to resolve real GitHub issues.

The release broke with OpenAI’s own recent practice: the company published no system card alongside GPT-4.1, the safety-evaluation document it had attached to every major model since GPT-4. A spokesperson said the model was “not a frontier model, so there won’t be a separate system card released for it,” and the company’s head of safety systems, Johannes Heidecke, argued it introduced no new modalities and did not exceed o3’s intelligence. Researchers in the AI safety community criticised the decision, arguing that a faster, cheaper, higher-performing model deserved more transparency about its risk profile, not less, and that the absence of a public evaluation made it harder for outside researchers to check OpenAI’s own judgement that the model was not worth the same scrutiny.

OpenAI brought GPT-4.1 and GPT-4.1 mini into ChatGPT itself about a month later, extending the undocumented models to Plus, Pro and Team subscribers and, for the mini version, to free users as well, though GPT-4o remained the default for the free tier. The episode became a reference point in later 2025 arguments about whether voluntary safety-disclosure commitments held up once a release did not look, by a lab’s own definition, dangerous enough to warrant one.