VISTA - Behavior: leaderboard
VISTA (2026, Visual Spec-To-App) tests coding agents on building and launching complete multi-page web apps from product specs, with or without screenshots and Figma structure, across 10 app tasks with 3,253 human-annotated interactions checked in a browser. This board is the mean behavior score from 0 to 1: whether each annotated interaction triggers the expected visible behavior; higher is better.
Metric: Mean behavior score (0-1): whether each annotated interaction triggers the expected application behavior, agent builds and launches a multi-page web app from the product spec for 10 tasks (128 pages, 3,253 annotated interactions), mean over the five prompt-information conditions C0 to C4 (C1 averaged over its three stacks), Playwright-driven interaction evaluator; higher is better. Source: arxiv.org. Saturation forecast: Around November 2027. 4 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | VISTA Claude Opus (checkpoint unspecified) | 0.336 | |
| 2 | GPT-5.4 | 0.33 | #135 |
| 3 | VISTA Claude Sonnet (checkpoint unspecified) | 0.313 | |
| 4 | GPT-5.5 | 0.283 | #54 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=vista-behavior · How It Works · Data refreshed daily, snapshot 2026-11-21.