// benchmark explainer
Terminal-Bench, explained
By the TrueIQ editorial team at Ahzy Capital · Last reviewed Oct 15, 2026 · How we work
Terminal-Bench puts an AI agent in a sandboxed Linux terminal and asks it to complete practical tasks, such as building software, fixing configs or processing data. Automated checks then verify the result.
| Made by | Stanford and the Laude Institute, with community contributors |
|---|---|
| Format | Agent works in a container; tests verify the end state |
| Versions | Several releases; Artificial Analysis runs a 'Hard' subset |
What it measures
It measures agentic skill: planning, running commands, reading output and recovering from errors over many steps. That's close to how coding agents are actually used.
Where it falls short
- Results depend a lot on the agent harness and time limits.
- Versions differ, so scores from different releases aren't comparable.
- Tasks are mostly developer and sysadmin work.
How to read a score
Check the version and the agent used. Compare models in the same harness.
Where to see current scores
We don't republish scores. Leaderboards change weekly, and numbers depend on how the test was run. Check these instead:
- Artificial Analysis: Terminal-Bench results (independently run)
- Terminal-Bench official site