Training the best computer use models in the world. Automate anything.
Today’s foundation models can generate words, but still cannot reliably operate a computer, where where most real economic value will be unlocked. We’re building models that learn from human behaviour at scale, powered by the largest dataset of computer interactions ever collected. In two months, we trained a model that beats GPT-5 and Claude 4 computer use at OSWorld, an industry standard computer use benchmark.