Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresVals' November 2026 revision of Vibe Code Bench 1–100: fifty held-out scenarios, each starting from a working Vibe Code Bench v1.1 web app, with up to ten dependent product requests (456 in all) and a shared ten-hour budget in a modified OpenHands v1 environment with Supabase, Stripe and MailHog. After each request a GPT-6 Luna browser agent runs new and regression workflows, and Claude Sonnet 5.5 reviews its failure verdicts: an iteration fails only on a confirmed implementation issue, and inconclusive reviews pass. Score is the mean fraction of planned iterations passed consecutively before the first failure. v1.1 rewrote v1's iteration prompts to be less ambiguous, audited every workflow assertion and added the failure review, so it is not comparable with vcb-1-100; fifty validation scenarios are excluded.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high · OpenHands | 49.7 | |
| 2 | GLM-5.3 | max · OpenHands | 29.8 | |
| 3 | MiMo-V2.6-Pro | not reported · OpenHands | 26.4 | |
| 4 | Grok 4.7 | x-high · OpenHands | 26.4 | |
| 5 | DeepSeek V4.1 Flash 1017 | max · OpenHands | 26.1 | |
| 6 | MiMo-V2.6-Flash | not reported · OpenHands | 24.5 | |
| 7 | MiniMax M3.1 Flash | max · OpenHands | 19.8 |