Reality check on computer-use agents: OSWorld 2.0 best system completes only 20.6% of long-horizon tasks

Medium (Adnan Masood)·medium signal

While computer-use agents climbed from 12% (OSWorld, April 2024) to ~85% success by mid-2026, the new OSWorld 2.0 benchmark — where the median task takes a human ~1.6 hours — sees the best frontier system finish just 20.6% of tasks. The gap exposes how far current GUI agents are from reliable long-horizon autonomy despite saturating older benchmarks. For builders it's a caution against extrapolating short-task demos into multi-hour workflow automation.