React Special Test: GPT-5.6 Sol Ranks First, But Best Score Only 43.1%

News on July 16: The Million team, specializing in React development tools, has launched ReactBench v1 to test whether AI programming agents can complete real-world React development tasks and avoid introducing new code issues. React is a mainstream JavaScript library used for building website and app interfaces. Previously, Million developed open-source tools such as React Scan, React Doctor, and Million.js. ReactBench selected 51 real tasks from open-source projects. In addition to checking whether functionalities work, it also uses over 400 rules to identify program errors, performance issues, accessibility problems, and code quality concerns. GPT-5.6 Sol ranked first with a comprehensive score of 43.1%, followed by Fable 5 at 41.2%. Officials stated that the gap between the two is small, and it is currently unclear if Sol is significantly stronger. When using the same XHigh configuration, the cost of a single test for Fable 5 is approximately 6.3 times that of Sol. Even with the highest-scoring configuration, less than half of the tasks were successfully completed. In 4,455 new feature development tests, the models collectively introduced 1,194 React issues, 77.5% of which were program errors or security problems.