Inference
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.5
Top resultBest result on this test.19.2%GPT-6 Astra
Top-3 spreadPoints from first to third.10.2 pts
UpdatedDate the source changed the results.18 Sept 2026

Result and cost

020406080100$100$300$1000$3000RESULTUS DOLLARS PER TASK · LOG SCALE

Results

5 results
#ModelResultThe score from the source.CostUS dollars to run one task.TimeTime to run one task.
01GPT-6 Astracodex19.2%$2,059.2832 s
02Claude Fable 5.1claude-code13.2%$595.9437 s
03Claude Opus 5claude-code9.0%$232.2325 s
04GPT-5.6 Solcodex8.3%$207.2519 s
05Gemini 3.8 Flashgoogle-computer-use4.2%$280.137 s