OSWorld-v2: code-writing agents lift GPT-6 Astra to 75.7
Alibaba's OpenGCUA has computer-use agents write each step as a program mixing GUI actions and shell commands; GPT-6 Astra rises from 67.1 to 75.7 with steps per task cut from 102.1 to 33.5.
Let a computer-use agent write Python instead of only clicking, and all eight frontier models tested get better. Mean partial score on OSWorld-v2 goes from 45.3 to 64.9, with 53% fewer steps per task. The Alibaba team behind OpenGCUA has the model write each step as one program that mixes GUI actions and shell commands. The program can check a result and pick the next move without another model call. The model also decides when it needs a screenshot and keeps its own todo list, two jobs the harness used to do by fixed rule. – GPT-6 Astra: 67.1 to 75.7, steps per task 102.1 to 33.5 – GLM-5.3-Flash: 29.6 to 64.9 – Code vs. plain GUI plus shell access: +6.1 points, 14.8% fewer steps The harness, prompts and trajectories from all eight models are released. Caveat: this is the 82-task offline set of OSWorld-v2. The real-device harness and cross-device SDK are early, tested on only four macOS tasks. https://arxiv.org/abs/2610.22200