GPT-6 Astra pushed "AI operating your computer" to the frontier, but the system behind it stays closed — outsiders watch demos and guess at the design. Alibaba's MAI-UI team offers a counter-move: OpenGCUA, an open-source project aiming to release the full stack for generalist computer-use agents (GCUAs). The first release ships with a technical report (arXiv:2610.22200, https://arxiv.org/abs/2610.22200) and an Apache-2.0 codebase under the Tongyi-MAI organization on GitHub.

Beyond screenshot-and-click

The team defines a GCUA as a system that decides for itself how to observe, act, and organize work: it reads the accessibility tree, calls app APIs, runs shell commands, and touches the GUI only when needed. The evaluation question shifts from "how well does the model drive a GUI" to "how well does the system finish the job by whatever means work".

Two core design moves

First, code execution: GUI actions are written as code and combined with API and shell calls in one program, where loops, branches, and intermediate results drive what happens next. Second, harness-to-model decision transfer: observation and task-state decisions move from fixed harness rules to the model — the harness supplies the mechanism, and the model decides when to use it, requesting a screenshot only when it needs one.

The numbers: all eight frontier models improve

On OSWorld-v2, eight models — including GPT-6 Astra, Claude Opus 5, and Gemini 3.8 Flash — rise from a mean partial score of 45.3 to 64.9 (+19.6), while steps per task fall 53% on average; GPT-6 Astra alone drops from 102 to 34 steps. Weaker models gain the most: GLM-5.3-Flash adds 35.2 points and DeepSeek-V4.1-Flash 29.8, shrinking the strongest-to-weakest gap from 39.1 to 18.0 points. An ablation shows that with identical GUI and shell access, writing operations as code beats a GUI+CLI agent by another 6.1 points and 14.8% fewer steps.

The cold water: a 19.4x token gap on real devices

The real-device side is the more telling part. With the harness held fixed across four macOS tasks, Qwen3.8-Max used 19.4 times the reasoning tokens of GPT-6 Astra and still completed fewer tasks. The team traces the gap to model behavior itself — exactly what their next release, an open GCUA model, targets.

So what

This release opens the implementation and prompts, evaluation trajectories from all eight models, a macOS real-device harness (a menu-bar app), and a cross-device SDK preview. For researchers, a frontier CUA design can finally be reproduced and improved systematically instead of reverse-engineered from demos. For the industry, the reminder is blunt: under the same harness, the real execution-efficiency gap between models is far larger than leaderboard gaps suggest — beyond "can it do the task", ask "how many tokens does it burn". Project: https://github.com/Tongyi-MAI/OpenGCUA