Learn
What is Terminal-Bench?
Last reviewed July 16, 2026
Terminal-Bench is a benchmark that drops an AI agent into a real command-line environment and scores whether it completes concrete tasks: installing and configuring software, debugging failing builds, processing files, wiring services together. Success is checked by the task's end state, not by judging the transcript. It measures the loop agents actually run (act, observe output, adapt), which makes it a direct signal for tool use.
Why the terminal is a good testbed
The shell gives exact, checkable feedback: commands fail loudly, state persists, and success is verifiable by inspecting the environment afterward. That forces the full agentic loop (plan, act, read output, correct course) rather than one-shot generation, and it resists partial credit: a half-configured system is a failure.
Reading Terminal-Bench scores
Scores reflect the model plus its agent harness: retry budgets, planning scaffolds, and tool wrappers all move results. Compare like with like, and treat a score as a floor for what a well-harnessed agent can do. Models strong here tend to be strong on long, multi-step work generally.
Top models on Terminal-Bench
Full leaderboardPeak scores from the compiled benchmark data; reasoning models use their highest effort tier.
idapt agents run real terminals on cloud computers and your paired machines, with the autonomy dial and budgets bounding what they may do: the same skill Terminal-Bench measures, with oversight.
Frequently asked
See it in practice
The fastest way to understand it is to use it. Start free, no credit card.
Try idapt free