Vals AI Vibe Code Bench: leaderboard
The model builds a working web application from a written specification, using over thirty tools to edit files, manage Supabase state, browse its running app and search the web. An autonomous browser agent then clicks through scripted user workflows against the deployed app. The number is the share of workflow tests that pass, and higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around January 2027. 114 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 | 92.39 | #11 |
| 2 | Gemini 4 Argon | 91.91 | #8 |
| 3 | Claude Haiku 5.5 | 90.44 | #40 |
| 4 | Claude Fable 5 | 90.35 | #16 |
| 5 | Claude Opus 5.5 | 90.29 | #3 |
| 6 | Claude Fable 5.1 | 90.26 | #9 |
| 7 | GPT-6 Astra | 89.59 | #6 |
| 8 | GPT-6.1 Sol | 88.93 | #7 |
| 9 | Claude Opus 5 | 88.4 | #12 |
| 10 | GPT-6 Sol | 87.82 | #14 |
| 11 | DeepSeek V4.1 Flash (1017) | 87.73 | #20 |
| 12 | Grok 4.7 | 86.17 | #25 |
| 13 | Muse Wonder Full | 86 | #10 |
| 14 | Muse Spark 1.3 (Max) | 85.86 | #24 (Muse Spark 1.3) |
| 15 | MiMo-V2.6-Pro | 85.22 | #42 |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-vibe-code-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.