Benchmarks · agentic
CUA-bench
Computer-use agents playing commercial video games through keyboard and mouse input, with half the titles held out.
8 models · 8 results · updated 2026-10-13 · vals.ai ↗
Leaderboard
GPT-6 Astra max
Heavy line: best score at each price. Thin lines join one model's effort levels. Cost uses each model's median subscription plan, or API price with no plan data. 3 more models under All models in settings.
More from this benchmark
1 published measuresTime per task9 results · lower is better
3 more models under All models in settings.
All results
8 rows| # | Model | Effort | CUA-bench | Subscription $ / task | API $ / task | Time | Harness |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra | max | 19.2% | $78 | $1821 | 32s | codex |
| 2 | Opus 5.5 | max | 14.0% | $2.85 | $188 | 15s | claude-code |
| 3 | Fable 5.1 | max | 13.2% | $50 | $324 | 37s | claude-code |
| 4 | Opus 5 | max | 9.0% | $4.03 | $232 | 25s | claude-code |
| 5 | GPT-5.6 Sol | max | 8.3% | $5.71 | $207 | 19s | codex |
| 6 | Muse Spark 1.3 | max | 5.8% | $72 | $143 | 36s | muse-code |
| 7 | Gemini 4 Argon | high | 4.8% | $138 | $194 | 71s | antigravity |
| 8 | Gemini 3.8 Flash | high | 4.2% | $11 | $280 | 37s | google-computer-use |