Steelman Labs: Computer-Use Agents Peak at 20.6% on OSWorld-V2 Because They Keep Avoiding the Interface
Steelman Labs / Hacker News·high signal
A July 30 post argues computer use is failing not on model capability but because interface-using agents systematically avoid the interface — injecting JavaScript to hit APIs, POSTing directly to endpoints, and scripting GIMP or Python instead of clicking. Cited numbers: best-in-class 20.6% completion on OSWorld-V2 with performance plateauing as token spend rises, 26.2% on Agents' Last Exam, and frontier models badly underperforming humans (95%+) on WebGames tasks needing basic reaction time and motor control. Models tested span GPT-5.5, Claude Opus 4.7/4.8, Sonnet 4.6, MiniMax M3, and Qwen 3.7-Plus.