ReactBench v1

ReactBench is a benchmark that evaluates coding agent performance on real-world-grounded React tasks. Unlike existing benchmarks, it doesn't just check whether tests pass — it also validates React performance, accessibility, and maintainability issues.

At the core of the benchmark is React Doctor, an open-source validator developed by a team of React experts. It uses over 400 rules to detect incorrect Effects, unnecessary re-renders, and accessibility problems. It was built based on tools actually used in production at GitHub, PayPal, Airbnb, and others.

The tasks being evaluated are extracted from merged PRs in real open-source projects, and agents perform two types of tasks: implementing new features or improving existing code.

Evaluation results: GPT-5.6 Sol (Medium) recorded the highest pass@1 at 43%, while Claude Fable 5 achieved 41.2%. The performance gap between leading models is minimal, but in terms of cost efficiency, GPT-5.6 Terra (Medium) is the optimal choice, delivering 38% performance at one-third the cost.

Issue analysis: Of the 1,194 React Doctor issues introduced across 4,455 Write React runs, 77.5% (925) were bugs. Common problems involved list rendering and Hook correctness. Write tasks tended to fail mainly on behavioral tests, while Fix tasks tended to fail on resolving React issues.

The benchmark was developed against the backdrop of real-world problems caused by model-generated code in production environments, such as Cloudflare's major outage in September 2025. Considerations included the fact that even a minimum 0.1-second speed improvement in semiconductors increases conversion rates by 8.4%, and that accessibility defects can create legal risk.