Terminal-Bench 4.0

Terminal-Bench 4.0 sits an AI model in front of a black-screen, text-only computer environment and grades how well it can find files, install programs, and fix problems. Think of it like handing a junior developer a real computer and saying 'go fix this bug,' then scoring them on the outcome.

This score matters because when companies release a new model, they don't just claim it's smarter than the last one — they back it up with this benchmark score. But the score isn't a pure measure of intelligence; it can also shift depending on how much a model's safety guardrails intervene.

In articles, you might see something like: "Anthropic explained that the gap between the two models' Terminal-Bench 4.0 scores came from the older, less refined cyber safety guardrails interfering with some tasks." The easy mistake here is thinking the score gap means the models differ in raw 'skill' — when actually it means the tightness of the safety guardrails placed on the same model got in the way of task performance.