Vals AI Launches CUA-Bench for Real-Time Computer Use

Vals AI has launched CUA-Bench, a benchmark that tests how well an AI model can play six games by working a keyboard and mouse in real time. Three of the six games are kept private. Vals AI chose computer use and games as examples of tasks that humans find easy and today's models still find hard.

The real-time requirement is the hard part the benchmark is built around. A game does not pause while a model decides what to press, so a model has to produce its actions quickly enough to keep up with what is happening on screen. Vals AI says that to saturate CUA-Bench, meaning to reach scores near the top of the scale, models will need to output actions in real time and learn continuously from video, not just from text. That is a statement about what the benchmark is designed to demand, not a result.

The launch post's text contains no scores, so it says nothing yet about how current models perform. It also does not name the games, explain how a run is scored, or say how the three private games are used. For now CUA-Bench is a new yardstick for a capability that Vals AI says today's models still find hard, and the evidence about any particular model will come when results are published.