ReactBench (39 tasks)
Code
Revised 39-task ReactBench panel. Five trials per task, passing both behavioral tests and the React Doctor no-new-issues gate. The publisher reclassified five tasks from fixing to writing by October 12, 2026, leaving this overall task cohort unchanged. Kept separate from the earlier 51-task panel.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 54.4 |
| Claude Fable 5.1 | 49.2 |
| GPT-5.6 Sol | 46.7 |
| Claude Fable 5 | 44.6 |
| Claude Opus 5 | 42.1 |
| GPT-5.6 Terra | 39.0 |
| DeepSeek V4.1 Flash | 35.4 |
| Gemini 3.8 Flash | 34.4 |
| GPT-5.6 Luna | 33.8 |
| Grok 4.6 | 33.8 |
| DeepSeek V4 Pro 0813 | 33.8 |
| Kimi K3 | 33.3 |
| GLM-5.3 Flash | 32.8 |
| Opus 4.8 | 32.3 |
| Sonnet 5 | 32.3 |
| Grok 4.5 | 32.3 |
| Grok 4.7 | 32.3 |
| GLM-5.3 | 30.8 |
| GLM-5.2 | 27.2 |
| Muse Spark 1.1 | 27.2 |
| Gemini 3.1 Pro Preview | 25.6 |
| Kimi K2.7 Code | 23.6 |
| Muse Spark 1.2 | 23.6 |
| Gemini 3.5 Flash | 23.1 |
| DeepSeek-V4-Flash-0731 | 21.5 |
| Composer 2.5 | 17.4 |
| Inkling-Small | 7.7 |