VA-Bench: leaderboard
VA-Bench (Zhang et al., 2026) tests whether general-purpose multimodal models can carry out simulated robot manipulation from visual demonstrations: the model watches an RGB-only expert video, chooses camera viewpoints, issues metric Cartesian arm commands for a fixed controller and revises them from execution feedback. Terminal task success in percent, macro-averaged over 14 task families (11 single-arm, 3 dual-arm) with 20 seeds each and averaged over three runs; higher is better.
Metric: Task success (%; all 14 base task families (11 single-arm, 3 dual-arm); 20 physically verified seeds per task family in RoboTwin simulation; the model watches an RGB-only expert video, writes its own task summary, chooses camera viewpoints and issues metric Cartesian arm commands that a fixed model-agnostic controller executes, with execution feedback; terminal success from the environment checker; task-family macro-average, mean of three runs). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max | 53.93 | #55 |
| 2 | Claude Opus 5 | 52.86 | #20 |
| 3 | GPT-5.6 Sol | 51.55 | #29 |
| 4 | GPT-5.6 Terra | 23.1 | #60 |
| 5 | Seed 2.1 Turbo | 15 | #156 |
| 6 | Gemini 3.6 Flash | 14.17 | #72 |
| 7 | GPT-5.6 Luna | 13.93 | #117 |
| 8 | Claude Sonnet 5 | 9.05 | #108 |
| 9 | Qwen 3.7 Max | 8.1 | #142 |
| 10 | Qwen 3.7 Plus | 7.14 | #203 |
| 11 | MiniMax-M3 | 4.05 | #265 |
| 12 | MiMo-V2.5 | 3.33 | #353 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=va-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.