OpenAI launches GPT-4.1 API family featuring 1M context and nano model

If you use AI to write code, you know the pain: it constantly rewrites entire files when you only asked for a three-line tweak, hallucinating changes you never wanted.

OpenAI just launched the GPT-4.1 family in the API, rolling out GPT-4.1, GPT-4.1 mini, and its first-ever GPT-4.1 nano. This release is built specifically for developers, bringing massive upgrades to coding, instruction following, and a 1M token context window across all three models. Its knowledge cutoff is refreshed to June 2024.

Alongside the launch, OpenAI announced it is deprecating the compute-heavy GPT-4.5 Preview, shutting it down on July 14, 2025, after giving developers three months to migrate.

Surgical coding and strict instruction following

On SWE-bench Verified, GPT-4.1 scored 54.6%, jumping 21.4 percentage points over GPT-4o and easily topping GPT-4.5. OpenAI CPO Kevin Weil pointed out on X that hitting 54 on this benchmark as a non-reasoning model is a huge leap.

Crucially, OpenAI tuned GPT-4.1 for real developer ergonomics. It more than doubled GPT-4o's score on Aider's polyglot diff benchmark. Instead of rewriting whole files, it reliably outputs unified diffs, reducing extraneous edits from 9% down to 2%. Max output tokens also doubled to 32,768 for those who still need full file generation.

Instruction following also saw a major bump, scoring 87.4% on IFEval and 38.3% on Scale's MultiChallenge. The model handles negative rules, ordered steps, and knowing when to say "I don't know" far more consistently, though early testers note it interprets prompts much more literally.

Aggressive price cuts and the first nano model

The pricing and latency shifts might be the most aggressive part of the rollout.

GPT-4.1 mini matches or beats GPT-4o across intelligence evals while cutting latency nearly in half and reducing costs by 83%.

The brand-new GPT-4.1 nano becomes OpenAI's cheapest and fastest model yet, carrying a 1M context window and scoring 80.1% on MMLU. Kevin Weil noted it costs just 12 cents per million blended tokens, making it primed for autocompletion and high-volume classification. GPT-4.1 itself is 26% cheaper than GPT-4o, and tools like Windsurf immediately rolled out 7 days of free access.

Why it matters

This release shows OpenAI pivoting hard toward real-world developer utility over vanity model sizes.

Sunsetting the ultra-heavy GPT-4.5 after just a short preview is a telling move. Enterprise builders do not want massive, slow compute monsters when a leaner model can perform better on coding and structured tasks at a fraction of the cost. We think killing off GPT-4.5 proves OpenAI is listening to what actually survives in production.

For engineers, the value is immediate. The combination of native diff adherence and 1M context removes the biggest bottleneck in automated code review and repo-level refactoring. You get faster loops, less bill shock, and far fewer broken builds.

In our view, the competitive pressure here shifts directly to rivals like Anthropic and Google at the API tier. When a mini model beats yesterday's flagship at an 83% discount and a nano model handles 1M context for pennies, the baseline expectation for everyday developer infrastructure has just reset.