OpenAI’s GPT-5.6 Sol posted the top score in a new evaluation that measures how well coding agents handle production-grade React changes. Running at extra-high reasoning effort, it solved 43.1 percent of tasks on the first attempt, averaged across five trials. A cheaper, medium-effort setting of the same model matched that score. The benchmark, called ReactBench, checks whether an agent’s code both functions correctly and avoids the sort of defects that have already knocked real products offline.
Million, the startup behind ReactBench, also built the open-source tools React Doctor, React Scan, and Million.js. Combined, those three projects have topped 53,000 GitHub stars, and teams at Airbnb, GitHub, PayPal, and Rippling have folded them into their own workflows. To build the benchmark, Million’s researchers started with 23,087 candidate pull requests drawn from public React codebases. Automated screening and a manual pass then whittled that pool down to 51 finished tasks. Each submitted solution has to clear a hidden functional test. It must also satisfy React Doctor’s library of more than 400 static-analysis rules, covering everything from buggy effects and wasted re-renders to accessibility gaps and long-term maintainability risk. Agents never see the grading logic directly: a separate, offline container checks the work only after the agent has finished.
Anthropic’s Fable 5 was the strongest non-OpenAI entry, reaching 41.2 percent at extra-high effort, less than two points behind Sol. It was also far pricier to operate: averaged across every effort level, Fable 5 cost $7.97 per trial against $1.37 for Sol, nearly six times as much. At their respective top efforts, it cost roughly 6.3 times what Sol did for a lower score. The rest of the field trailed further behind. Anthropic’s Opus 4.8 reached 34.1 percent at its highest setting, and Z.ai (formerly branded Zhipu AI) hit 32.9 percent with GLM 5.2. Anthropic’s Sonnet 5 reached 30.6 percent, and Kimi K2.7 Code, built by China’s Moonshot AI, finished last among the eight model families tested, at 20.4 percent.
Code quality did not track cost the way one might expect. Over 4,455 write-task trials, ReactBench’s graders logged 1,194 new issues flagged by React Doctor, and outright bugs made up 925 of those, or 77.5 percent. List-rendering mistakes and incorrect Hook usage showed up most often. Sol and Fable 5 kept their bug counts lowest, at roughly 15 per 100 trials apiece. Kimi K2.7 Code racked up more new issues than any other family: 67.4 per 100 trials, over three times Sol’s rate.
The two task types break down differently. Write React asks an agent to build a real feature drawn from a merged pull request, while Fix React asks it to clean up an existing component without breaking what already works. Of the 2,486 write attempts that failed, 65.3 percent produced code clean enough to pass React Doctor but still missed the hidden functional test. In other words, the agent wrote tidy code that did the wrong thing. Of the 3,219 fix attempts that failed, 60.8 percent passed the functional test while still tripping React Doctor’s checks. In that case, the agent patched the immediate bug without touching the pattern that caused it.
ReactBench frames the stakes with three outside data points. Cloudflare has said a September 2025 outage in its dashboard and API traced back to a single flawed useEffect dependency that slipped past human review and testing. Separately, Google has measured a mobile page-speed gain of just one-tenth of a second lifting retail conversion by 8.4 percent. ReactBench also points to a 27 percent jump last year in federal lawsuits over web accessibility in the United States. None of those figures came from ReactBench’s own testing, and the writeup does not claim its tasks caused any of them. It uses them instead to argue that React-specific bugs carry a measurable cost, a case few competing agent benchmarks bother to make.
The results are self-published by the benchmark’s own creators, not an independent lab. ReactBench’s own writeup admits the leading configurations are bunched too tightly to declare a confident winner. Even the top scorer still loses more often than it wins, clearing under half of all attempts. For any team routing production React changes through autonomous agents, that ceiling matters more than the leaderboard order. A pass@1 rate near 43 percent means most agent-written pull requests still need a human review pass before merge. Right now, the cheaper OpenAI configuration beats the priciest Anthropic one on both cost and code quality.
ReactBench published these findings in its “ReactBench v1” writeup on the project’s own blog at reactbench.com, dated July 15, 2026.
