On Monday, OpenAI unveiled a new suite of models under the GPT-4.1 name. Yes, “4.1” — adding another layer to the company’s already complex naming system.
The lineup includes GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, all of which, according to OpenAI, excel in coding and following instructions. These multimodal models are accessible via OpenAI’s API — not ChatGPT — and feature a 1-million-token context window, allowing them to process around 750,000 words at once (more than the length of War and Peace).
GPT-4.1 launches amid rising competition from companies like Google and Anthropic, who are accelerating efforts to develop advanced programming models. Google’s recently released Gemini 2.5 Pro, which also supports a 1-million-token context window, performs strongly on popular coding benchmarks. Anthropic’s Claude 3.7 Sonnet and Chinese AI startup DeepSeek’s upgraded V3 models also rank highly.
Many tech giants, including OpenAI, aim to train AI coding models that can tackle complex software engineering tasks. OpenAI’s broader vision is to create an “agentic software engineer,” as described by CFO Sarah Friar at a tech summit in London last month. The company envisions future models that can build complete apps from start to finish — including quality assurance, bug testing, and writing documentation.
GPT-4.1 marks a step toward that goal.
“We’ve optimized GPT-4.1 for real-world use based on direct developer feedback — focusing on areas they care about most: frontend coding, reducing unnecessary edits, consistent formatting, structured responses, tool usage, and more,” an OpenAI spokesperson told. “These improvements enable developers to build agents significantly better at real-world software engineering.”
OpenAI states that GPT-4.1 outperforms its GPT-4o and GPT-4o mini models on coding benchmarks like SWE-bench. While GPT-4.1 mini and nano offer higher speed and efficiency, they trade off some accuracy. Notably, GPT-4.1 nano is OpenAI’s fastest and most cost-effective model to date.
Pricing is as follows: GPT-4.1 costs $2 per million input tokens and $8 per million output tokens. GPT-4.1 mini is priced at $0.40 per million input tokens and $1.60 per million output tokens, while GPT-4.1 nano is just $0.10 per million input tokens and $0.40 per million output tokens.
Based on OpenAI’s internal tests, GPT-4.1, which can output more tokens at once than GPT-4o (32,768 vs. 16,384), achieved between 52% and 54.6% on SWE-bench Verified — a human-validated subset of SWE-bench. (OpenAI noted that some solutions couldn’t run on its infrastructure, which accounts for the score range.) These results fall slightly below those reported by Google’s Gemini 2.5 Pro (63.8%) and Anthropic’s Claude 3.7 Sonnet (62.3%) on the same benchmark.
In another evaluation, OpenAI tested GPT-4.1 with Video-MME, which assesses how well a model understands video content. GPT-4.1 topped the charts with 72% accuracy in the “long, no subtitles” video category, according to OpenAI.
While GPT-4.1 performs decently on benchmarks and benefits from a more recent knowledge cutoff (June 2024), it’s worth noting that even today’s best models still struggle with tasks experts handle with ease. Research has shown that code-generating models often fail to fix — or may even introduce — bugs and security vulnerabilities.
OpenAI also acknowledges that GPT-4.1’s reliability drops as input length increases. In one of the company’s own tests (OpenAI-MRCR), the model’s accuracy fell from around 84% at 8,000 tokens to just 50% at 1 million tokens. Additionally, GPT-4.1 tends to be more “literal” than GPT-4o, often requiring more explicit, detailed prompts.
