Vals AI CUA-bench: leaderboard
Computer-use agents playing six commercial video games with keyboard and mouse only, under a latency budget, with three of the six held out. The headline accuracy is the mean of the six games, each scored out of 100; the board's separate public and hidden columns exist to expose overfitting rather than to rank. Higher is better. Source: vals.ai/benchmarks/cua_bench.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 19.17 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 14 | #3 (Claude Opus 5.5) |
| 3 | Claude Fable 5.1 (Max) | 13.17 | #10 (Claude Fable 5.1) |
| 4 | Claude Opus 5 (Max) | 9 | #16 (Claude Opus 5) |
| 5 | GPT-5.6 Sol (Max) | 8.33 | #22 (GPT-5.6 Sol) |
| 6 | Muse Wonder (Max) | 7.17 | #18 (Muse Wonder) |
| 7 | Muse Spark 1.3 (Max) | 5.83 | #31 (Muse Spark 1.3) |
| 8 | Gemini 4 Argon (High) | 4.83 | #9 (Gemini 4 Argon) |
| 9 | Gemini 3.8 Flash (High) | 4.17 | #30 (Gemini 3.8 Flash) |
No result here: #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-cua-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.