EU and US could reach trade deal this weekend - Reuters
Investing.com -- OpenAI has unveiled the GPT-4.1 family of models—GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano—marking a significant leap in performance over previous releases like GPT-4o. These models are designed to excel in coding, instruction following, and long-context understanding, with support for up to 1 million tokens.
OpenAI’s CEO, Sam Altman, stated, "Benchmarks are strong, but we focused on real-world utility, and developers seem very happy."
GPT-4.1 delivers a 54.6% score on SWE-bench Verified, a 21.4% improvement over GPT-4o, making it a top choice for software engineering tasks. It also scores 38.3% on Scale’s MultiChallenge and 72.0% on Video-MME, setting new benchmarks in instruction following and multimodal comprehension.
Performance gains are paired with efficiency: GPT-4.1 mini cuts latency by nearly half and slashes costs by 83% compared to GPT-4o. Despite its smaller size, it frequently outperforms GPT-4o in both speed and quality.
GPT-4.1 nano, the most compact model, still delivers strong results with an 80.1% score on MMLU, while maintaining full support for a 1 million token context window. It’s particularly well-suited for fast, cost-effective tasks like classification, autocompletion, and powering AI agents.
The models are tuned for real-world utility, incorporating feedback from developers to improve reliability in practical applications. Companies like Windsurf and Qodo report up to 60% gains in coding efficiency and code review quality.
GPT-4.1 also shines in legal and financial contexts, with Thomson Reuters (NYSE:TRI) seeing a 17% boost in multi-document review accuracy and Carlyle reporting a 50% improvement in financial data extraction. Its new multi-hop reasoning capabilities, tested on the Graphwalks dataset, enable complex workflows like legal and technical document cross-referencing.
Vision capabilities have been enhanced as well—GPT-4.1 mini achieves 75% on the MMMU image benchmark, surpassing GPT-4o. Long-context video processing is another highlight, with a 6.7% jump over GPT-4o on the Video-MME benchmark.
GPT-4.1 reportedly beats GPT-4.5 in several key benchmarks, and in turn, OpenAI announced they will discontinue GPT-4.5 in the OpenAI API over the next three months, citing GPU concerns. Kevin Weil, OpenAI’s CPO, called GPT-4.5 "a very successful experiment."
All GPT-4.1 models are available via API and come with competitive pricing, starting at $0.10/$0.40 per million tokens for nano. With GPT-4.5 Preview set to sunset by July 14, 2025, GPT-4.1 offers a more powerful and cost-efficient solution for developers and enterprises alike.