which AI gets the job done

An agent is an AI you give a whole task to on a computer — fix this, build that, tidy these files — and it works through it on its own, running steps and checking its own results. That is a different skill from answering a question well, which is why a model can top a chat ranking and sit mid-table here.

Not ours. The data is arena.ai’s agent leaderboard, built from over a million real working sessions, and it is linked on every view. We add no scoring of our own — we put it into plain words, and say plainly when our copy has gone stale rather than quietly serving old numbers.

Every number here is “how much better or worse than the average model”, so a minus figure means below average rather than bad. Each one also carries a margin of error, and two models sitting a fraction apart have not been shown to differ. Where the page says “too close to call”, believe it.

A ranking built from a million sessions moves slowly, so polling hourly would spend courtesy on nothing. The data refreshes once a week, a scheduled check re-tests the connection every Monday, and this page states its own age so a broken refresh cannot go unnoticed.