Salesforce's StateAct cuts computer-use costs 9x, lifts accuracy

Salesforce AI Research's new computer-use agent hits 26.9% success on OSWorld 2.0, roughly 9x cheaper per task than the same model reading screenshots, per a paper posted to arXiv.

Computer-use agents get billed and benchmarked on cost per completed task, and vision-based inference (reading a full screenshot after every click) is the expensive part. A 9x cost cut, with partial success up from 54.8% to 61.6%, changes that math for anyone running these agents at scale.

The agent, called StateAct, does most of its work by reading and editing files, backends, and the DOM directly through code instead of screenshots. It calls in a screenshot-reading GUI subagent only when code access isn't available.

That fallback fired on just 28 of 108 test tasks, covering 1.1% of the agent's steps.

The 20.6% figure it beat belongs to a screenshot-only version of the same Opus 4.8 model StateAct runs on.

Yes, but a code-only variant with no GUI subagent reached just 45.9% partial success, below the screenshot baseline, so the fallback earns its keep.