What does a Terminal-Bench score mean?

A Terminal-Bench score averages seven very different task categories, software, science, ML, operations, hardware, security, and media, and that average can hide near-total failure in some of them. On the 4.0 leaderboard, the top model, Claude Opus 5.5, leads at 61.62% overall, but that number comes from strength that isn’t evenly spread: it posts 78.79% on ML and only 48.15% on operations, its weakest category among the ones it still leads. Hardware and media are the harder story for the field as a whole: seventeen of the thirty-one models scored zero on each, while operations and hardware post the lowest average scores across every model tested, 11.7% and 12.5%.

A high overall score, in other words, says a model handles the categories it’s strong in, not that it can do all seven kinds of work. The same gap shows up in Gaia2’s own scores, where one split lands near 80% and a harder one lands at zero on the identical run. Check which categories actually drove a Terminal-Bench number before reading it as one verdict on an agent’s terminal skills.