ReactBench leaderboard exposes coding-agent gaps

ReactBench is a stronger production-readiness signal than test-only coding benchmarks, but its results still show frontier agents are far from turnkey.

• –Agents must modify real open-source React codebases rather than solve synthetic prompts
• –React Doctor catches regressions that ordinary unit tests miss, including broken effects and performance issues
• –Opus 5.5 leads, yet its roughly 54% pass rate means nearly half of first attempts still fail
• –Rankings vary significantly by model, effort level, harness, and rollout cost
• –The benchmark does not fully measure visual correctness or proprietary application architectures