OpenAI's recent launch of the GPT-4.1 model family, including mini and nano versions, marks a significant upgrade over previous iterations. With a refreshed knowledge cutoff of June 2024, these models are designed to excel in coding and instruction following, achieving up to 54.6% on SWE-bench Verified and a 38.3% score on Scale's MultiChallenge benchmark. A major focus has been placed on real-world usability, while also enhancing latency and cost efficiency, notably reducing costs by 83% for the GPT-4.1 mini model, making them suitable for various applications.
The new GPT-4.1 models outperform previous versions significantly, especially in coding and instruction following, with enhanced context comprehension capabilities.
GPT-4.1 leads in coding with a 54.6% score on SWE-bench Verified and shows a remarkable improvement of over 21% from GPT-4o.
The models significantly improve latency and cost efficiency, with GPT-4.1 mini reducing both latency by nearly half and cost by 83%.
Real-world utility was a priority in the training of these models, ensuring they meet developer needs effectively.
Collection
[
|
...
]