Best agents to use computers
Long-horizon, real-world desktop workflows. Partial score is the standard: it measures how far an agent gets through work that takes a person over an hour and a half.
Ranked by partial score
From the official OSWorld 2.0 leaderboard, full task set. Partial score is our headline metric because it credits real progress on long tasks that few agents finish outright.
| # | Model |
|---|
Rows mix two releases (v2.1 and v2026.08.08). Compare scores within the same release.
Reproduce links and trajectories are the ones each team submitted to the official board. Source: osworld-v2.xlang.ai, updated November 14, 2026.
Benchmarks for computer-use agents
The evaluations used to measure agents that see a screen and act with a mouse, keyboard or touch.
No benchmarks match.
Best computers for agents
Hosted machines with a full desktop that an agent can see and control. Compared on operating system, what survives between sessions, price and how agents connect.
| Platform | OS | Desktop | Persistence | Pay as you go | Plans | Free to start | Agent access |
|---|
† From third-party comparisons; not confirmed on the vendor's own pricing page.
Hourly rates are for different machine sizes, so compare them with the size shown. Prices are in USD from vendor pricing pages, checked November 18, 2026.
Not listed: Scrapybara, which has moved on to a different product, and code sandboxes without a desktop (Modal, Vercel Sandbox, Fly.io Sprites).