OpenAI has announced the arrival of three new models: GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, which promise to take artificial intelligence to a new level. With context windows of up to 1 million tokens and a knowledge cut-off extending to June 2024, these models are designed to surpass their predecessors.
According to the company, GPT-4.1 and its variants outperform the recently updated GPT-4o, offering exclusive API performance at this time, so it will not be integrated into ChatGPT immediately. The performance data is impressive: a 21.4% increase in the SWE-bench Verified score and a record in understanding long videos.
Performance and Speed Improvements
OpenAI has not only focused on accuracy but also on speed. The GPT-4.1 model returns its first token in approximately 15 seconds when processing 128,000 tokens, and up to 30 seconds for a million. The mini and nano versions are even faster, with responses in under 5 seconds for similar inputs.
For real-world use cases, OpenAI has collaborated with partners like Thomson Reuters, which reported a 17% increase in the accuracy of legal document review when using GPT-4.1. On the other hand, Carlyle found that the new model is 50% more effective in extracting data from complex documents.
Visual intelligence has also improved significantly. In various vision benchmarks, GPT-4.1 mini has shown superior performance compared to its predecessor across multiple tasks. This means that image understanding is now more advanced, expanding possibilities for developers and businesses.
In terms of costs, using prompt caching and the Batch API can further reduce expenses, making it attractive for large-scale applications. With the withdrawal of the GPT-4.5 Preview scheduled for July 2025, OpenAI is focused on delivering more robust and cost-effective performance through GPT-4.1.
0 Comments