OpenAI has introduced a series of GPT-4.1 AI models that specialize in programming

Мирослав Трінько - 15 April, 09:42 AM

OpenAI has announced a new series of artificial intelligence models — GPT-4.1, which includes three versions: full, GPT-4.1 mini and GPT-4.1 nano. These models are not available in ChatGPT, but can be used through the company's API. All of them have a context window of 1 million tokens, which allows you to work with text volumes of up to 750 thousand words.

OpenAI says the new models are particularly well-suited for programming and following instructions. The company says GPT-4.1 has been improved based on developer feedback. It writes better code for websites, makes fewer unnecessary changes, is more accurate in the format of responses, and uses the right tools more reliably.

GPT-4.1 competes with models from Google (Gemini 2.5 Pro) and Anthropic (Claude 3.7 Sonnet), which also have millions of context windows and demonstrate high results in benchmarks. GPT-4.1 showed up to 54.6% accuracy in the verified SWE-bench Verified test, which is slightly lower than Gemini 2.5 Pro (63.8%) and Claude 3.7 Sonnet (62.3%). In the video content test (Video-MME), GPT-4.1 achieved 72% accuracy in the difficult category "long videos without subtitles".

The mini and nano models are cheaper and faster, although less accurate. Regarding cost:

GPT-4.1: $2 per million input tokens and $8 per output;
GPT-4.1 mini: $0.40 / $1.60;
GPT-4.1 nano: $0.10 / $0.40 is the cheapest OpenAI model.

Although GPT-4.1 performs well, it has its drawbacks. In particular, its accuracy decreases when processing very long queries. In addition, the model often takes instructions too literally and requires clear and specific instructions. At the same time, OpenAI sees GPT-4.1 as an important step towards creating an AI agent capable of independently creating full-fledged software products.