ProductWebBench - Stage Success: leaderboard
ProductWebBench: frozen frontend product changes with incremental milestones, 12 turns per unit and cumulative browser gate replay. Up to 400 language and 400 vision attempts; Table 1 coverage varies from 385 to 400 and text-only models lack vision results. The available-task mean uses each model’s actual coverage. Whole-task success requires all stages and final gates. FP8, Preview and Code identities remain distinct. Rate (%).
Metric: Rate (%). Source: arxiv.org. Saturation forecast: Around 2031. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 42.45 | #30 |
| 2 | Claude Opus 4.8 | 41.69 | #37 |
| 3 | GLM-5.2 FP8 | 39.9 | #196 |
| 4 | Kimi K3 | 39.3 | #33 |
| 5 | DeepSeek V4 Flash | 38.24 | #144 |
| 6 | Kimi K2.7 Code | 38.1 | #92 |
| 7 | DeepSeek V4 Pro Preview | 37.82 | #140 |
| 8 | Qwen 3.7 Max | 37.32 | #79 |
| 9 | MiniMax-M3 | 36.42 | #151 |
| 10 | Gemini 3.5 Flash | 34.88 | #63 |
| 11 | Qwen 3.6 35B A3B | 33.46 | #232 |
| 12 | Claude Sonnet 4.6 | 32.65 | #94 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=productwebbench-stage-success · How It Works · Data refreshed daily, snapshot 2026-10-25.