OpenAI Launches GPT-4.1 via API, Retires GPT-4.5

OpenAI today unveiled three new models available exclusively through its API: GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano — the latter marking the company's first-ever "nano" model. OpenAI says all three outperform GPT-4o and GPT-4o mini across the board, with notable gains in coding and instruction-following, and that they support context windows of up to one million tokens. All three carry a refreshed knowledge cutoff of June 2024.

The announcement comes alongside a decision that signals where things are headed: OpenAI will begin deprecating GPT-4.5 Preview in the API, shutting it down in three months, on July 14, 2025. The reasoning is straightforward — GPT-4.1 delivers equal or better performance on many key capabilities at a much lower cost and latency.

What's new in GPT-4.1

OpenAI's central pitch is coding. On SWE-bench Verified, a benchmark that measures real-world software engineering skills, GPT-4.1 solves 54.6% of tasks versus 33.2% for GPT-4o — a 21.4-point improvement over GPT-4o and 26.6 points over GPT-4.5. OpenAI itself qualifies the figure: its measurement excludes 23 of the 500 problems whose solutions couldn't run on its infrastructure; counting those as failures brings the score down to 52.1%.

On instruction-following, GPT-4.1 scores 38.3% on Scale's MultiChallenge benchmark, 10.5 points above GPT-4o. And on multimodal long-context comprehension, it sets a new record on Video-MME, scoring 72.0% in the long-video, no-subtitles category — 6.7 points higher than GPT-4o.

Beyond the numbers, OpenAI stresses that it trained these models with real-world usefulness in mind, working closely with its developer community. Two specific improvements reflect that goal. First, GPT-4.1 follows diff formats — the format that expresses only the lines that change in a file, rather than rewriting the whole thing — more reliably, saving on cost and latency. On Aider's polyglot benchmark, it more than doubles GPT-4o's score and beats GPT-4.5 by 8 points. Second, in OpenAI's internal evaluations, unnecessary code edits dropped from 9% with GPT-4o to 2% with GPT-4.1.

For those who prefer rewriting entire files, the output limit rises to 32,768 tokens, double GPT-4o's 16,384. And in frontend coding — the visible part of a web application — human evaluators preferred websites generated by GPT-4.1 over those from GPT-4o 80% of the time.

Mini and nano: betting on price and speed

The structural novelty here is that the family spans the entire latency curve. According to OpenAI, GPT-4.1 mini matches or exceeds GPT-4o in intelligence while cutting latency by nearly half and cost by 83%. That's a considerable leap for a small model that, in several tests, outperforms the previous generation's flagship.

GPT-4.1 nano is the fastest and cheapest model in the lineup. It keeps the one-million-token context window and scores 80.1% on MMLU, 50.3% on GPQA, and 9.8% on Aider's multilingual coding benchmark, ahead of GPT-4o mini. OpenAI positions it for tasks like classification or autocomplete, where speed matters more than sophistication.

This tiering of sizes and prices answers a practical problem: not every task needs the most powerful model, and paying for capacity you don't use is a drag on any product trying to scale. Offering three tiers lets developers pick the exact balance of cost, speed, and intelligence they need.

Agents: the real target

The improvements in instruction reliability and long-context comprehension are aimed squarely at agents — systems capable of completing tasks autonomously on a user's behalf. OpenAI says that, combined with tools like its Responses API, these models make it possible to build more useful agents for software engineering, extracting information from lengthy documents, or resolving customer queries with minimal supervision.

This is where the million-token context stops being a spec-sheet flourish. An agent working across an entire code repository or mountains of documentation needs to retain and actually use that information — a huge window is of little use if the model gets lost inside it. OpenAI argues that GPT-4.1 doesn't just have more context; it makes better use of it.

The company backs the launch with testimonials from its alpha testers. Windsurf says GPT-4.1 scores 60% higher than GPT-4o on its internal coding benchmark, that its users found it 30% more efficient at using tools, and roughly 50% less prone to repeating unnecessary edits. Qodo, after pitting GPT-4.1 against other leading models on 200 real GitHub pull requests under identical conditions, says it produced the best suggestion 55% of the time.

API-only, and a farewell

One nuance worth flagging: GPT-4.1 will only be available through the API, not in ChatGPT. OpenAI explains that many of its improvements in instruction-following, coding, and intelligence have already been gradually rolled into the latest version of GPT-4o within ChatGPT, and that it will keep porting more of them over in future releases. The split is telling: the company is running its consumer product on one track and its developer catalog on another.

The retirement of GPT-4.5 Preview fits that same logic. OpenAI introduced it as a research preview to explore a large, compute-intensive model, and says it learned a great deal from developer feedback. It's now shutting it down because GPT-4.1 does the same job, or better, for far less money and with lower latency. The company promises to carry the creativity, writing quality, humor, and nuance that users valued in GPT-4.5 forward into future API models.

The overall message is efficiency over raw power. A large, expensive model is being retired in favor of a family that's cheaper, faster, and — according to OpenAI's own figures — just as capable, if not more so, at the tasks that matter to developers. For them, the three-month window until July 14 is the deadline to migrate their applications before GPT-4.5 stops responding.