StateAct: Program State Cuts Agent Costs by 9x, Sets New SOTA

_akhaliq · x · 2026-07-28

While current computer-use agents rely heavily on vision models to read screen pixels, StateAct proposes a counterintuitive approach: using program state (such as files, backends, and DOM) as the primary interface, only invoking a GUI specialist for the 1.1% of steps that truly require visual judgment.

This "look less" approach sets a new SOTA on the OSWorld 2.0 benchmark. Using the exact same Claude Opus 4.8, the method increased task success rates from 20.6% to 26.9% while drastically reducing token consumption (from 224K to 100K), slashing cost-per-task from $72 to $7.8.

Furthermore, the framework is model-agnostic; dropping in an internal 31B model (SFR-CUA) yielded significant leaps across multiple benchmarks.

Original post →

More from coding & agent

coding & agent channel →