OpenAI has released three new AI models, GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. The models are now available through an API. According to OpenAI, the new models surpass older models in coding and understanding instructions.
These models handle up to 1 million tokens, allowing them to process vast amounts of information. They also understand long texts better, with updated knowledge from June 2024.
GPT-4.1 shines in coding, scoring 54.6% on a test called SWE-bench Verified, which checks real-world software skills. This is much higher than older models. It also excels in following instructions, scoring 38.3% on a benchmark named MultiChallenge, which tests multi-step tasks. For long videos or texts, GPT-4.1 scores 72.0% on Video-MME, a test for understanding extended content without subtitles.
Real-world applications and efficiency
Developers helped shape these models for practical use. GPT-4.1 mini outperforms some larger models while cutting response time by nearly half and costs by 83%. GPT-4.1 nano, the fastest and cheapest, handles tasks like sorting or auto-filling text efficiently, scoring 80.1% on MMLU, a knowledge test. These models also power agents, systems that work independently, like analyzing documents or coding software.
Only available via the API, GPT-4.1’s improvements appear in the latest ChatGPT version. An older model, GPT-4.5 Preview, will stop working on July 14, 2025, as GPT-4.1 offers better results at lower costs. OpenAI expects developers to find GPT-4.1 more reliable for coding tasks. It also creates better web apps, with 80% of human testers preferring its designs.
In real-world tests, companies like Windsurf saw 60% better coding results, while Qodo noted improved code reviews. Blue J improved tax research accuracy by 53%, and Thomson Reuters boosted document review by 17%. These models handle large datasets, like financial records, with 50% better accuracy. With faster response times and lower costs, these models could help developers build smarter, more efficient systems.