OPENAI announced today that he is publishing a new family of artificial intelligence The optimized models to excel in coding, as it increases efforts to push the increasingly rigid competition of companies like Google and Anthropic. The models are available for developers via the OPENAI (API) application programming interface.
OPENAI publishes three model sizes: GPT 4.1, GPT 4.1 Mini and GPT 4.1 NANO. Kevin Weil, product manager at Openai, said on a live flow that the new models are better than the most used model of Openai, GPT-4O and better than his largest and powerful model, GPT-4.5, in some respects.
GPT-4.1 marked 55% on Swe-Bench, a widely used reference to assess the prowess of coding models. The score is of several percentage points above that of other Openai models. The new models are “excellent in coding, they are excellent in the complex education that follows, they are fantastic for the construction of agents,” said Weil.
The ability of AI models to write and modify the code has improved considerably in recent months, allowing more automated means of software prototyping and improving supposedly so-called capacities AI agents. Rivals like Anthropic And Google both introduced models that are particularly good for writing code.
The arrival of GPT-4.1 has been largely rumoring for weeks. Openai apparently tested the model on certain popular rankings under the pseudonym Alpha Quasar, according to sources. Some users of the “stealth” model reported Impressive coding capacities. “Quasar solved all the open problems that I had with another genié code (sic) via the LLM which was incomplete,” wrote a person on Reddit.
All new models can analyze eight times more code at a time, which improves their ability to make improvements and correct bugs. The new models are also better to follow the instructions given by users, reducing the need to repeat orders in different ways to obtain the desired result. Openai has shown demos of GPT-4.1 by creating different applications, including a flash cards application for language learning.
“The developers care a lot about coding, and we have improved the ability of our model to write functional code,” said Michelle Pokrass, who works on post-training in Openai, on Monday Livestream. “We have worked to have it followed different formats and to better explore standards, run unit tests and write compilation code.”
GPT-4.1 is 40% faster than GPT.4O, the most used OpenAi model for developers. The cost of user entry requests has been reduced by 80% in this latest version, explains Openai.
On Livestream today, Varun Mohan, CEO of Windsurf, a popular tool for AI coding, said that the company had tested GPT-4.1 and found that the new model was “better at 60%” than GPT-4O according to its own benchmarks. “We have found that GPT-4.1 has much less degenerate behavior,” said Mohan, noting that the new model spends less time reading and editing unrelevant files by mistake.
In the past two years, Openai has aroused feverish interest in CatA remarkable chatbot First unveiled at the end of 2022In a growing company selling access to more advanced chatbots and AI models. In a TED interview last week, Altman said that Optai had 500 million weekly active users and that the use “increased very quickly”.
