ViBench 2.0
Code
Replit's ViBench 2.0 measures vibe coding: from an empty project, a coding agent builds one of 17 real-world web apps (8 public, 9 held out) one feature at a time from short product specs, choosing its own stack, in Replit's builder harness on the OpenHands SDK. A grading agent uses each finished app in a browser and follows its test plans step by step. The score is pass@1, the share of app builds in which every test plan passes, averaged over four builds per app and then over apps.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 64.7 |
| GPT-6 Astra | 60.3 |
| Claude Opus 5.5 | 48.5 |
| Claude Sonnet 5.5 | 19.1 |
| Grok 4.7 | 1.5 |
| DeepSeek V4.1 Flash | 0.0 |
| GLM-5.3 | 0.0 |
| Kimi K3 | 0.0 |
| Gemini 3.8 Flash | 0.0 |