Million, the team focused on React tooling, has released ReactBench v1, a benchmark for AI coding agents working on real React development tasks. The first result set shows that even the top-performing model, GPT-5.6 Sol, reached only a 43.1% success rate. None of the tested models crossed 50%, leaving a sizable gap between current AI performance and full replacement of human developers in production-grade React work.
Million is known for React Scan and Million.js. With ReactBench v1, the team moved beyond the older pattern of checking whether an AI system can pass basic unit tests. The new benchmark places models into real open-source project settings and checks whether the generated code meets production standards.
51 tasks drawn from real open-source projects
ReactBench v1 selected 51 tasks from real open-source repositories, framed as Pull Requests. The benchmark focuses on two core capabilities: Write React and Fix React.
To catch code that appears workable but introduces hidden problems, the test uses a validator called React Doctor. It includes more than 400 rules designed to flag invalid renders, poor performance, accessibility shortcomings, and maintainability issues. Million said the requirement is not just to finish the feature. The agent also has to avoid adding any new code errors.
GPT-5.6 Sol ranks first, with Claude Fable 5 close behind
Based on the published average pass rate, or pass@1, the leading AI models still have substantial room to improve. OpenAI’s GPT-5.6 Sol, in Medium / XHigh configurations, posted a combined score of 43.1% to take first place. Anthropic’s Claude Fable 5, in XHigh, followed with 41.2%.
Million said the gap between the two is narrow, and the current results are not enough to establish Sol as the undisputed leader.
No model cleared the 50% threshold
That headline result runs through the entire benchmark: every tested model finished below 50% success. The data also shows that across 4,455 new feature development tests, the models introduced 1,194 React-related issues in total. Of those, 77.5% were classified as severe bugs or security problems.
The report said list rendering and correct handling of Hook rules were the areas where models were most likely to fail.
Cost matters, and mid-tier models may be attractive
The report pointed to cost as another important variable. GPT-5.6 Terra, in Medium, scored 38.0%, putting it fairly close to the top tier on performance. At the same time, when both were compared under XHigh settings, Fable 5’s per-test cost was about 6.3 times that of Sol.
For companies that need code generation at scale, that spread suggests mid-tier models may offer a stronger cost-performance tradeoff.
React’s market share raises the stakes
Million said React accounts for about 70% of websites that use JavaScript frameworks. The team warned that if developers trust AI-generated React code without enough scrutiny, even small defects can be amplified in production, with possible consequences including system outages, lower conversion rates, and revenue losses.
Million said ReactBench is intended to give developers a more realistic way to measure whether AI agents can actually produce React code that is safe, compliant, and efficient.