ReactBench (39 tasks): Fixing React (11 tasks)
Code
October 2026 classification of the unchanged 39-task ReactBench panel: 11 refactoring tasks requiring target React Doctor findings to be removed without regressions, five trials per task. Kept separate from the historical 16-task fixing subset after five tasks moved to Write React.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 27.3 |
| GPT-5.6 Terra | 21.8 |
| Claude Opus 5 | 20.0 |
| Gemini 3.8 Flash | 20.0 |
| Gemini 3.5 Flash | 20.0 |
| Claude Opus 5.5 | 20.0 |
| Grok 4.6 | 18.2 |
| Claude Fable 5.1 | 16.4 |
| DeepSeek V4 Pro 0813 | 16.4 |
| DeepSeek V4.1 Flash | 16.4 |
| GPT-5.6 Luna | 14.5 |
| Grok 4.5 | 14.5 |
| Muse Spark 1.1 | 14.5 |
| Grok 4.7 | 14.5 |
| Gemini 3.1 Pro Preview | 12.7 |
| Muse Spark 1.2 | 12.7 |
| GLM-5.3 | 10.9 |
| Claude Fable 5 | 7.3 |
| Kimi K3 | 7.3 |
| Kimi K2.7 Code | 7.3 |
| Opus 4.8 | 5.5 |
| GLM-5.2 | 5.5 |
| DeepSeek-V4-Flash-0731 | 5.5 |
| GLM-5.3 Flash | 3.6 |
| Sonnet 5 | 3.6 |
| Composer 2.5 | 1.8 |
| Inkling-Small | 1.8 |