OpenAI has introduced GPT-4.1, a new family of multimodal AI models aimed squarely at developers. The launch includes GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, all available through OpenAI's API but not ChatGPT.
The company says the models “excel” at coding and instruction following. That positioning matters because AI labs are increasingly competing on whether their systems can handle real software engineering work, not just short coding prompts.
A coding model family for API users
GPT-4.1 is not a single model release. OpenAI is offering three versions: GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. The full model is presented as the strongest option, while mini and nano trade some accuracy for speed and efficiency.
All three models have a 1-million-token context window. In practical terms, the source describes that as roughly 750,000 words in one go, longer than “War and Peace”. For developers, the point is not only scale; it is the ability to provide a model with much more surrounding material before asking it to reason, edit, or generate.
OpenAI says the models are tuned for the parts of software work that developers notice quickly: frontend coding, following formats reliably, keeping response structure and ordering, using tools consistently, and making fewer unnecessary changes. Those details suggest the company is trying to make the models more useful inside developer workflows, where precision and predictable output matter.
The race to build stronger programming models
GPT-4.1 arrives as OpenAI faces pressure from other AI companies working on sophisticated programming systems. The source points to Google, Anthropic, and Chinese AI startup DeepSeek as competitors that have recently pushed coding performance forward.
Google's Gemini 2.5 Pro also has a 1-million-token context window and ranks highly on popular coding benchmarks. Anthropic's Claude 3.7 Sonnet and DeepSeek's upgraded V3 are also described as strong performers in this area.
The bigger goal is broader than code snippets. OpenAI CFO Sarah Friar described the company's ambition as an “agentic software engineer” during a tech summit in London last month. OpenAI says future models will be able to program entire apps end-to-end, including quality assurance, bug testing, and documentation writing.
GPT-4.1 is framed as a step toward that goal. The release is not presented as a finished autonomous engineer, but as a move toward models that can handle more of the ordinary complexity of software development.
Benchmarks, pricing, and model tradeoffs
OpenAI says GPT-4.1 outperforms GPT-4o and GPT-4o mini on coding benchmarks, including SWE-bench. The model can also generate more tokens at once than GPT-4o: 32,768 versus 16,384.
On SWE-bench Verified, a human-validated subset of SWE-bench, OpenAI's internal testing put GPT-4.1 between 52% and 54.6%. OpenAI noted that some solutions to SWE-bench Verified problems could not run on its infrastructure, which is why the result is given as a range.
Those results are below the figures reported by Google and Anthropic for the same benchmark: Gemini 2.5 Pro at 63.8% and Claude 3.7 Sonnet at 62.3%. That makes GPT-4.1 competitive, but not the top reported performer in the comparison described by the source.
Pricing is split by model and by input versus output tokens:
- GPT-4.1 costs $2 per million input tokens and $8 per million output tokens.
- GPT-4.1 mini costs $0.40/million input tokens and $1.60/million output tokens.
- GPT-4.1 nano costs $0.10/million input tokens and $0.40/million output tokens.
OpenAI says GPT-4.1 nano is its speediest and cheapest model ever. That gives developers a cheaper path for tasks where speed and cost matter more than maximum accuracy.
Long context is powerful, but not a cure-all
GPT-4.1 also performed strongly in a video-focused evaluation. OpenAI tested it with Video-MME, which is designed to measure whether a model can “understand” video content. The company says GPT-4.1 reached 72% accuracy in the “long, no subtitles” video category.
The model also has a more recent “knowledge cutoff” than some earlier systems, up to June 2024. That gives it a newer frame of reference for current events, according to the source.
Still, the launch comes with important limits. The source notes that even strong code-generating models can struggle with tasks that experts would handle, and that studies have shown such models may fail to fix security vulnerabilities and bugs or even introduce them.
OpenAI also acknowledges that GPT-4.1 becomes less reliable as the input grows. On OpenAI-MRCR, one of the company's own tests, accuracy fell from around 84% with 8,000 tokens to 50% with 1 million tokens.
That is the central tension of the release. A 1-million-token context window gives developers far more room to provide material, but more context does not automatically mean better answers. OpenAI also says GPT-4.1 can be more “literal” than GPT-4o, which may require more specific and explicit prompts.
What developers should take from GPT-4.1
For developers, GPT-4.1 is best understood as an API-first coding release with several distinct choices. The full GPT-4.1 model targets stronger coding performance. GPT-4.1 mini and GPT-4.1 nano are positioned around speed, efficiency, and lower cost.
The model family also shows where OpenAI wants to move next: tools that can follow structure, work with long inputs, and support more realistic software engineering tasks. But the source makes clear that benchmark gains and larger context windows do not remove the need for careful review, testing, and explicit instructions.
GPT-4.1 may be a meaningful upgrade for developers building agents and coding tools through the OpenAI API. It is also another sign that the competition around AI coding models is becoming one of the main fronts in the broader AI race.