Alibaba's OpenGCUA cuts OSWorld-V2 steps per task 53 percent across eight models

Counting the steps a desktop agent takes to finish a task is the clearest way to tell whether a design change made the agent better or only used more tokens. OpenGCUA, published by Alibaba's Tongyi MAI-UI team, counted those steps across eight frontier models.

The repository is Apache-2.0 and holds an OSWorld integration example with reference agents, a preliminary macOS real-device harness, a cross-device SDK for computers, browsers and phones, a native macOS menu-bar app, and the full evaluation trajectories from the eight models the paper tested. Weights and training code are not included. The repo has 212 stars.

The number

On the OSWorld-V2 offline evaluation set the paper reports a mean partial score rising from 45.3 to 64.9, a 19.6 point gain, and a 53 percent average cut in steps per task across GPT-6 Astra, Claude Opus 5, Gemini 3.8 Flash and five others. GPT-6 Astra went from 102 steps per task to 34. The method is written up in a technical report at arXiv 2610.22200, with ablations and case studies.

The gain comes from the harness design, not a trained policy. OpenGCUA lets the agent decide at each step what to observe (a screenshot, an accessibility tree, a shell response), how to act (a GUI click, a scripted command, a code run), and how to organise the work, instead of following fixed rules about which to use when. Combining the four interaction modes under one agent is the "generalist" in the name.

The weak spot in the measurement

Everything on this page is Alibaba's own evaluation of its own stack. OSWorld-V2 is a public benchmark, so an outside team can reproduce the setup, but no independent group has yet reported the same numbers. The eight models are also run inside OpenGCUA, meaning the baseline for each one is that model under a reference agent, which is lower than the strongest tuned configuration a vendor ships. The paper names the baselines, so a buyer can compare, but a reader who sees "GPT-6 Astra went from 102 to 34" without that context will overstate the result.

The step-count metric is the one worth borrowing from this release even if the headline numbers do not survive outside testing. A team running a computer-use agent today usually tracks wall-clock time and token spend, neither of which separates "the agent got lucky and finished early" from "the agent took five tries to click the right button and the first four do not show up in the cost line". Steps per task is cheap to log, repeatable across runs, and the one metric that moves when the agent gets better at choosing what to try.

Source

OpenGCUA: an open-source stack for generalist computer-use agents, Tongyi-MAI, and the OpenGCUA tech report, Alibaba Tongyi MAI-UI team.