Inference
EN
Vals AI

CUA-bench

Can an AI agent play six commercial video games with only a keyboard and a mouse? Half of them are held out.

ModelsModels with a result.9
Top resultBest result on this benchmark.19.2%GPT-6 Astra
Top-3 spreadPoints from first to third.6.0 pts
UpdatedDate the source changed the results.13 Oct 2026

Result and cost

9 of 9 results

Results

9 results
#ModelResultThe score from the source.CostUS dollars to run one task.TimeTime to run one task.
01GPT-6 Astracodex19.2%$1,821.3732 s
02Claude Opus 5.5claude-code14.0%$188.0415 s
03Claude Fable 5.1claude-code13.2%$323.7837 s
04Claude Opus 5claude-code9.0%$232.2325 s
05GPT-5.6 Solcodex8.3%$206.9919 s
06Muse Wondermuse-code7.2%$597.2131 s
07Muse Spark 1.3 Maxmuse-code5.8%$143.4836 s
08Gemini 4 Argonantigravity4.8%$193.7871 s
09Gemini 3.8 Flashgoogle-computer-use4.2%$280.0337 s