OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use
libraryYour notes
From the same MAI-UI team at Alibaba Token Hub that shipped MAI-Gym the day before, and the counterpart to it: MAI-Gym is where a computer-use agent is trained and measured, OpenGCUA is the stack it runs in. A generalist computer-use agent is defined here as "a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces" — the emphasis being on chooses.
Two design claims. Code execution is treated as central rather than auxiliary: it "lets intermediate results determine subsequent operations and enables models to apply their existing programming capabilities directly to computer interaction", so the agent computes over what it sees instead of only clicking. And harness-to-model decision transfer, "moving selected observations and task-state decisions from fixed harness rules to the model" — which is a direct position in the argument Model–Harness Interactions tracks: if capability is a property of the pairing, one response is to move the decisions out of the harness entirely.
The headline is measured across eight models including GPT-6 Astra and Claude Opus 5, which makes it a stack result rather than a model result: on OSWorld-v2 the full configuration improves task scores on all eight, raising mean partial score by +19.6 points while cutting reported steps per task by 53.0% against batched GUI-only execution. More scores for fewer steps is the combination worth noting — scaffolds usually buy one by spending the other.
One incidental measurement deserves repeating because almost nobody reports it: across four real macOS tasks, selected Qwen3.8-Max runs used 19.4× as many reasoning tokens as GPT-6 Astra. Alibaba publishing that about its own flagship, in its own paper, is a more useful disclosure than the benchmark table. Released as the implementation and prompts, evaluation trajectories from all eight models, and preliminary versions of a real-device harness and a cross-device SDK; code Apache-2.0, paper CC BY 4.0.