OpenGCUA: Alibaba's open computer-use stack cuts steps 53%
Report, trajectories, macOS harness and SDK all public, tested across eight frontier models. Also today: RTX Spark laptops ship; Accenture caps Claude spend.
Copy markdown
One policy for clicks, shell, API and code
OpenGCUA unifies GUI actions, shell commands, API calls and code execution in a single agent policy. Across eight frontier models — GPT-4 Astra, Claude Opus 5 and Gemini 3.8 Flash among them — OSWorld-v2 scores rose 19.6 points on average and steps per task fell 53% versus GUI-only runs. The paper, evaluation trajectories, a macOS harness and a cross-device SDK are all out — enough to wire a desktop automation agent that does more per step.
RTX Spark laptops hit shelves, from $2,599
The first N1X-based RTX Spark laptops ship today from Microsoft, ASUS, Dell, Lenovo and MSI, starting at $2,599 (Surface Laptop Ultra, ASUS ProArt P14) and topping out at $3,799 for the Dell XPS 16; HP's two arrive Nov 4. Surface configs reach 128GB unified memory — enough to keep sizeable open-weight models resident locally instead of renting a cloud GPU.
Accenture caps company-funded Claude at $29 a head
Per an internal memo reported by 404 Media, Accenture will centrally fund only $29 per employee per month of Claude usage; staff who burn through it must find project budget or lose access. A blunt reminder that at scale, per-seat AI spend is now a metered line item — worth knowing if your workflow assumes an open tap at work.