What is GPT-4.1, with its mini and nano versions, and what is it good at?

Yesterday, Monday, April 14, 2025, OpenAI launched a new family of models for developers: a full-size GPT-4.1 plus two smaller versions, mini and nano, the latter being the first "nano" model OpenAI has offered. If you use ChatGPT, you will not find them in the model menu. They are only available through OpenAI's API, the service that apps and programs use to send requests to OpenAI's models.

The name invites confusion, since GPT-4.5 came out in February. GPT-4.1 is not a step back, though. It is a practical model tuned for the things developers asked for most: writing and editing code, following instructions exactly, and handling very long inputs, at lower prices.

What OpenAI released

OpenAI says all three do better than their predecessors, GPT-4o and its mini version, and all three share the same headline features: a context window of up to one million tokens, up from 128,000 for GPT-4o, and knowledge of the world up to June 2024. (A token is a piece of a word. TechCrunch puts a million tokens at roughly 750,000 words, more than War and Peace.) GPT-4.1 can also write longer replies, up to 32,768 tokens instead of 16,384.

Here is the lineup with launch prices, per million tokens:

OpenAI says GPT-4.1 costs 26% less than GPT-4o for a typical request, and that mini matches or beats GPT-4o on many tests while cutting response time nearly in half and cost by 83%. Long inputs cost nothing extra beyond the normal per-token price, and the discount for prompt caching, reusing the same opening text across requests, rises to 75%.

OpenAI also announced that GPT-4.5 Preview will be switched off in the API on July 14, 2025, because GPT-4.1 offers similar or better results on many tasks for far less money. In ChatGPT, OpenAI says, many of the same improvements are being folded into GPT-4o instead.

What improved

Coding. On SWE-bench Verified, a test built from real bugs in open-source software, GPT-4.1 solves 54.6% of tasks, compared with 33.2% for GPT-4o. OpenAI notes that it left out 23 of the 500 problems that would not run on its systems; counting those as failures gives 52.1%. TechCrunch points out that Google and Anthropic report higher scores on the same test: 63.8% for Gemini 2.5 Pro and 62.3% for Claude 3.7 Sonnet. OpenAI also says GPT-4.1 makes far fewer unnecessary edits to code, 2% against 9% for GPT-4o, and is better at producing a diff, a list of just the changed lines, instead of rewriting whole files.

Following instructions. GPT-4.1 scores 87.4% on IFEval, a test of instructions that can be checked automatically (such as "answer in under 100 words"), against 81.0% for GPT-4o. It also improved on keeping track of instructions given earlier in a long conversation. One side effect: testers find it takes prompts more literally, so vague requests get literal answers.

Long inputs. OpenAI hid a single fact, a "needle", at different points in very long texts, and all three models found it every time, right up to a million tokens. Harder tests show the limits. On OpenAI's MRCR test, where the model must pick out one of several near-identical requests buried in a long conversation, accuracy falls from about 84% with 8,000 tokens to 50% at a million, TechCrunch reports. On long videos without subtitles, GPT-4.1 scored 72.0% on the Video-MME test, ahead of GPT-4o's 65.3%.

Why it matters

GPT-4.1 is less about a dramatic leap than about making useful models cheaper and more dependable for builders. Better instruction following and long-context reading are exactly what agents need, programs that carry out multi-step tasks on a user's behalf, and OpenAI suggests combining the models with its new Responses API to build them. Early testers report concrete gains: Thomson Reuters says its legal assistant reviewed multiple documents 17% more accurately, and the coding tool Windsurf measured 60% better results than with GPT-4o on its internal benchmark.

The competition is clear in the background. Google's Gemini 2.5 Pro already offers a million-token window, and Anthropic's Claude 3.7 Sonnet ranks highly for coding. GPT-4.1 is OpenAI's answer for developers who want strong coding at a lower price.

Which GPT-4.1 model should you choose for your project?

Start from the job, not the name. Choose GPT-4.1 for work where mistakes are costly or the input is large: editing code, reviewing contracts, answering questions across a big set of documents, or powering an agent that takes several steps. Choose mini for most everyday features, such as drafting replies, summarizing tickets or extracting data from forms; OpenAI's results suggest it often matches GPT-4o for much less. Choose nano when speed and volume matter most and the task is simple, such as tagging messages by topic or suggesting the next word.

A sensible method is to build a small test set of 20 to 30 real examples from your own work, run them through two models, and compare quality, speed and cost before deciding. Because GPT-4.1 follows instructions closely, be explicit:

You sort customer emails. Reply with exactly one label from this list:
billing, shipping, returns, other.
Reply with the label only, in lower case, with no explanation.

A prompt like this suits nano well. If the labels start drifting on tricky emails, move up to mini, and keep the full model for the tasks that truly need it.