OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use

OpenGCUA. Toward a Fully Open Stack for Generalist Computer Use. By Hanzhang Zhou and fourteen others, from Alibaba Group. Published on the fifteenth of October. Its code repository is worked out: about one hundred and eighty source files, with a readme and a licence.

Computer use has reached a turning point, the authors argue. The focus is widening from interaction through the graphical interface alone to systems that complete tasks on a user's computer through both visual and programmatic interfaces. They give this kind of system a name, a generalist computer-use agent, or GCUA: one in which the model itself chooses how to observe, how to act, and how to organize its work across those interfaces. Frontier implementations exist, GPT-6 Astra's computer-use system among them. They remain largely closed, and that limits any systematic study of how they are designed. A design that cannot be inspected cannot be compared, ablated, or built upon.

Open Generalist CUA, or OpenGCUA, is an ongoing project to develop the full stack for these agents and to open it, with the community. This report is its first release. Two design choices sit at the centre. The first is code execution, a key feature of the agent. It lets intermediate results determine the operations that follow, and it lets models apply the programming capabilities they already have directly to computer interaction. The second is what the authors call harness-to-model decision transfer. Selected observations and task-state decisions are moved out of fixed harness rules and handed to the model.

Both choices push in the same direction: less is decided in advance by the surrounding scaffold, and more is decided by the model while the task runs. The release around them is broad. It includes the implementation and its prompts, and evaluation trajectories from all eight models that were tested. It also includes preliminary versions of two further pieces. A real-device harness addresses challenges on actual machines that benchmarks do not fully capture. A cross-device SDK provides a common programming interface across computers, mobiles, and browsers.

On OSWorld-v2, the full configuration improves task scores for all eight evaluated models, GPT-6 Astra and Claude Opus 5 among them. Measured against batched GUI-only execution, mean partial score rises by nineteen point six points, and reported steps per task fall by fifty-three percent on average. Higher scores and fewer steps arrive together, rather than one bought at the cost of the other.

A second observation comes from the real-device harness. Across four real macOS tasks, selected Qwen3.8-Max runs used nineteen point four times as many reported reasoning tokens as GPT-6 Astra. The authors read this as a substantial difference between models in how efficiently they execute on real devices, a gap that a benchmark score on its own would not show.

The case being made is that openness comes before understanding. With closed frontier systems, the central design questions stay out of view: how much of the work goes through code rather than the visual interface, and which decisions belong to the harness rather than the model. OpenGCUA puts those questions into a stack others can run. Its first answers are that code execution and handing decisions to the model raise scores for every model tried, and that real devices expose differences in efficiency that only show up when the agent runs on an actual machine.