Alibaba's OpenGCUA Lifts 8 Models by 19.6 Points on OSWorld-v2

Across eight frontier models, Alibaba's Tongyi-MAI team reports a +19.6-point mean partial-score lift on OSWorld-v2 when each one runs inside its new OpenGCUA stack, with reported steps per task falling 53.0% versus a batched GUI-only baseline. The paper says the gap between the strongest and weakest evaluated model 'shrinks by more than half.'

OpenGCUA is what the authors call a generalist computer-use agent, defined in the paper as 'a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces.' GUI clicks, accessibility calls, API calls, and shell commands all live inside one program with loops, branches, and intermediate results. The second design move is 'moving selected observations and task-state decisions from fixed harness rules to the model,' which the paper labels harness-to-model decision transfer.

The evaluated models include GPT-6 Astra, Claude Opus 5, and Qwen3.8-Max.

There is a cost side. 'Selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra,' the paper notes, on a set of four real macOS tasks. The sample is too small to generalize, but Tongyi-MAI published the figure anyway.

The release bundles a macOS menu-bar harness for real-device execution, a cross-device SDK spanning computer, browser, and phone, and the full evaluation trajectories from all eight models. It is the third Tongyi/Alibaba write-up on our tracker this week, after Waypoint-OPD on distillation gating.