ReactBench
2026-10-10 · Ecosystem eval · v1.2
- What it measures
- Write features and fix issues in real React repositories. v1.2 has 39 tasks, 37 repositories, and 76 configurations. Million runs it. The React team does not, and neither does Meta.
- How
- The v1.2 board in the public reactbench.com/data file. Each effort tier is its own row. Tiers of one model are not merged into one score. This page lists the 47 rows that map to models on this site. The rank is the rank in the full 76-row board.
- What the score means
- The percent is the pass rate. The interval is the confidence interval in the data file. Cost is the average spend of that configuration.
- Caveats
- This is an ecosystem eval, not an official React or Meta eval. Do not call it official, and do not merge it with the Supabase, Next.js, or Convex evals into one number. The data file has no separate run date. The version is v1.2. Configurations without a model on this site are left off this table.
- Updates
- Synced v1.2 of reactbench.com/data on 2026-10-10.
- Source
- reactbench.com/data
Baseline board
| # | Model | Baseline | Cost per run |
|---|---|---|---|
| 1 | Claude Fable 5.1 max | 49.2 | $5.50 |
| 2 | GPT 5.6 Sol max | 46.7 | $3.06 |
| 3 | Claude Fable 5.1 xhigh | 46.2 | $3.74 |
| 4 | Claude Fable 5 max | 44.6 | $12.76 |
| 5 | Claude Fable 5.1 high | 44.1 | $1.74 |
| 6 | Claude Fable 5 xhigh | 43.1 | $9.17 |
| 7 | Claude Opus 5 max | 42.1 | $5.94 |
| 8 | Claude Opus 5 high | 41.5 | $3.68 |
| 9 | GPT 5.6 Sol xhigh | 41.0 | $2.02 |
| 10 | Claude Fable 5 high | 40.5 | $6.80 |
| 11 | GPT 5.6 Terra max | 39.0 | $1.13 |
| 12 | GPT 5.6 Terra high | 37.4 | $0.40 |
| 13 | GPT 5.6 Sol high | 36.9 | $1.58 |
| 14 | GPT 5.6 Terra xhigh | 36.9 | $0.60 |
| 15 | Qwen3.8 Max xhigh | 36.4 | $1.62 |
| 16 | Claude Opus 5 xhigh | 36.4 | $4.93 |
| 18 | Claude Fable 5 low | 35.4 | $3.05 |
| 19 | DeepSeek V4.1 Flash max | 35.4 | $0.27 |
| 20 | GPT 5.6 Sol medium | 34.9 | $1.10 |
| 22 | Claude Opus 5 medium | 34.4 | $2.04 |