Glossary
Task success rate
The share of tasks an agent finishes under a named scoring rule, on a named task set, with the date and the step cap attached.
A task success rate is a fraction. The numerator is the tasks the agent finished. The denominator is the task set. The scoring rule, the step cap, the judge, and the date belong in the same sentence. Drop them and two vendors can print numbers that do not compare.
OSWorld 2.0, read November 13, 2026, prints 20.6 percent binary completion for Claude Opus 4.8 at a 500-step cap on 108 long tasks, and a 54.8 percent partial score beside it. Binary and partial are not substitutes. The Online-Mind2Web leaderboard file, read the same day, carries submitted averages above 90 percent on 300 live-web tasks, next to a March 2025 row at 30.0 percent for an older Browser Use run. Those files are not one rate.
A benchmark is the frozen set plus the rule. A behavioral evaluation checks one action rather than the suite. The harness is part of what the rate measures. How browser agents produce the rate, and which product pages to read before you buy one, is How do AI browser agents complete tasks?. The web-page tool under test is browser use.