Is computer use reliable yet?

Computer-use agents now complete most short desktop tasks but still fail most long professional workflows. The OSWorld 2.0 paper from XLANG Lab at the University of Hong Kong reports frontier agents at 79 to 83% on the original OSWorld, and the best agent on its multi-hour workflows, Anthropic’s Claude Opus 4.8, at 20.6%.

Newer models have raised that number without closing the gap. The highest official score on the full task set of the OSWorld 2.0 leaderboard, as of its October 15, 2026 update, is Anthropic’s Claude Opus 5 at 44.33% on the OSWorld 2.1 release, still under half.

The paper ties the failures to task length: Anthropic’s Claude Opus 4.7 and OpenAI’s GPT-5.5 complete 20 to 24% of tasks a person finishes in under 45 minutes, no model passes 10% at 137 to 163 minutes, and completion hits zero past that, the same decline agents show on longer tasks generally.

The paper finds agents lose track of constraints, miss information that arrives mid-task, guess instead of asking the user, and skip verification. Each model it profiled spends under 7% of its budget on recovery and repair, so earlier mistakes go largely unrepaired, and an agent that skips verification can report a task done that isn’t. Those behavior profiles come from one run per model, so the authors treat them as descriptive, not causal.