OpenAI releases GPT-4.1, mini and nano in the API

Background

OpenAI ships its models through two different doors. The API is the door for software: other companies' programs call the model directly. ChatGPT is the door for people: a chat product with its own release schedule. A model can come out through one door before the other, and what it is tuned to do can differ between the two.

The GPT-4.1 launch was one of those split releases. The reference points OpenAI measured the new family against were the models already in service: GPT-4o, GPT-4o mini and GPT-4.5 ^[4].

The release

On 14 April 2025 OpenAI launched three models at once in its API — GPT-4.1, GPT-4.1 mini and GPT-4.1 nano ^[1]. The three differ in size, a distinction that matters to developers because a smaller model is normally cheaper and faster to run, at the cost of capability on hard tasks.

At launch, OpenAI said the flagship GPT-4.1 would be available only through the API, and that its instruction-following and coding gains would instead be folded into later versions of GPT-4o inside ChatGPT ^[4]. The model did later arrive in ChatGPT for paid users, on 14 May 2025 ^[1].

The same day brought a removal as well as an addition. OpenAI deprecated the GPT-4.5 preview in the API on 14 April 2025 ^[3], and said it would be turned off on 14 July 2025 ^[4].

Key details

The most concrete technical number attached to the release was the context window: about 1 million tokens, given precisely as 1,047,576 ^[2]. A token is a small chunk of text, often a word or part of a word, and the context window is the total amount of text a model can take into account at once — the instructions, the documents, the earlier turns of a conversation and the answer it is writing. A window of a million tokens is large enough that a developer can hand over long documents, or a lot of code, in a single request instead of splitting them into pieces and stitching the results together.

On quality, OpenAI said all three models beat GPT-4o and GPT-4o mini across the board, with the biggest gains in coding and in following instructions ^[4].

The headline benchmark was SWE-bench Verified, where GPT-4.1 scored 54.6% ^[4]. That benchmark poses real software issues and checks whether the model's patch makes the failing tests pass, so it measures work done rather than recall. The score was 21.4 points above GPT-4o and 26.6 points above GPT-4.5 ^[4].

Why it matters

The release showed how fast OpenAI's model line was turning over. GPT-4.5 preview was deprecated in the API on the day GPT-4.1 arrived ^[3], with a shut-off date of 14 July 2025 ^[4], so anyone building on it had to plan a migration.

The naming pointed the same way. GPT-4.1 sits inside the GPT-4 line rather than starting a new generation, and OpenAI did not give it a separate ChatGPT debut: the same improvements were pushed into later GPT-4o versions instead ^[4]. The company also shipped one family in three sizes rather than a single model, which lets developers trade cost against capability depending on the task; all three supported a context window of about 1 million tokens ^[2]. For people building products on the API, that combination — a large window, a stated coding jump and a small cheap variant — was the practical content of the release.