Claude's Computer Use Scores 83.5% — Then Drops to 20.6% on Real Work
A new benchmark stretched computer-use tasks from a few clicks to 1.6-hour workflows. The best model's completion rate collapsed by a factor of four. The gap is the whole story of production.
NeuroX AI · July 16, 2026
The headline number for computer-use agents keeps climbing: Claude Opus 4.8 reaches 83.5% on OSWorld-Verified, the standard desktop-agent benchmark. Then OSWorld 2.0 lengthened the tasks to real workflows — a median of 1.6 hours of human work each — and the same model dropped to 20.6% binary completion, roughly one task in five.
Read the two numbers together. On short, narrow tasks — click here, fill that field — the agent looks nearly solved. On the 108 long-horizon tasks that actually resemble a workday, spanning multiple apps and hours of maintained state, it finishes fewer than a quarter of them. Partial-score accuracy is 54.8%, which means it usually gets started and rarely gets done.
This is the prototype-to-production gap with a clean measurement on it. The demo — the short task — is where the 83.5% lives. The job is the long horizon, where errors compound, state drifts, and one wrong step 40 minutes in poisons everything after it. No model release closes that gap for you; the jump from Opus 4.7 to 4.8 moved binary completion 18.2% → 20.6%.
What closes it is the layer around the model: checkpoints, recovery, verification at each step. That layer is code you own, not a benchmark you wait for.