The leaderboard is current and actively updated, so this ranking reflects roughly where things stand now.
OSWorld
OSWorld tests whether an AI agent can actually operate a real computer: open real apps, click around, edit files, and chain steps across programs to finish a task, versus just describing how to do it.
Tasks run on a real operating system and are graded by checking the end state, so the agent has to genuinely get the job done.
An agent is given a real desktop task, for example 'in this spreadsheet, sort the rows by date and save it', and must drive the actual application to completion. A script then checks the resulting file, so a near-miss that leaves the file wrong scores zero.
This is the closest thing to a can-it-use-my-laptop-for-me test, so it is the key signal for anyone weighing an agent for real desktop or web automation. It is also a striking progress story: at the 2024 launch the best AI finished only about 12 percent of tasks versus 72 percent for humans, and by mid-2026 the top agents have passed the human number.
Success rate on real computer tasks (the 361-task verified split), scored by per-task execution checks, higher is better. Maintained by the OSWorld team; the verified board is current as of mid-2026.
Humans complete about 72 percent of these tasks. At the November 2024 launch the best AI managed only 12.24 percent; by mid-2026 the leaders reach the low-to-mid 80s, so the field has gone from far below human to above the human reference in under two years.
The biggest catch is what a model means here: the top entries are full agent scaffolds (for example Pointer Agent with Opus 4.7), not the bare model you would call from an API, and the same scaffold appears twice with different base models, so the ranking partly measures the harness, not the model. The top three are also a near-tie (within about two points). The official page is JavaScript-rendered; the verified data lives in a downloadable spreadsheet.
The benchmark creates the need to know. The catalog explains the ideas behind it:
The top ranks are a statistical tie. Read this as a cluster, not a clean number one.