App-Bench
Agents · 2025-10-25
App-Bench evaluates hosted app builders and coding assistants on six full-stack web-application tasks covering healthcare, real estate, finance, legal services, and education. Each tool receives three independent one-shot attempts per app under its default Pro settings. The best attempt for each app is selected, and the aggregate divides completed feature checks by all 151 manually verified binary rubric checks across the six apps. Six counts app tasks, not binary score outcomes; the headline is a rubric aggregate with max-of-runs selection.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.5 | 67.5 |
| Gemini 3 Pro Preview | 50.3 |
| GPT-5.1-Codex-Max | 38.4 |
| Composer 1 | 27.8 |
| Gemini 2.5 Pro | 0.0 |