VA-Bench: leaderboard
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 December 2026. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.8 Max | 53.93 |
| 2 | Claude Opus 5 | 52.86 |
| 3 | GPT-5.6 Sol | 51.55 |
| 4 | GPT-5.6 Terra | 23.1 |
| 5 | Seed 2.1 Turbo | 15 |
| 6 | Gemini 3.6 Flash | 14.17 |
| 7 | GPT-5.6 Luna | 13.93 |
| 8 | Claude Sonnet 5 | 9.05 |
| 9 | Qwen 3.7 Max | 8.1 |
| 10 | Qwen 3.7 Plus | 7.14 |
| 11 | MiniMax-M3 | 4.05 |
| 12 | MiMo-V2.5 | 3.33 |
Interactive version: theaggregate.ai/benchmark?slug=va-bench · How It Works · Data refreshed daily, snapshot 2026-09-26.