Terminal-Bench
Open benchmark of agent tasks in terminal environments, with a public leaderboard of models and coding agents.
0 members list Terminal-Bench, unchanged over the last 12 weeks, data as of 18 November 2026
Members who list Terminal-Bench in their Stack, by week.
Members who list Terminal-Bench
Be the first to add Terminal-Bench
Nobody on Stackness runs Terminal-Bench yet. Add it to your stack and you're the first member running it.
About Terminal-Bench
Terminal-Bench is an open benchmark designed to evaluate agents and models in terminal environments. It measures how well coding agents and language models can handle tasks that require command-line interaction and execution. The benchmark provides a public leaderboard where developers can compare the performance of different models and agents across standardized tasks, tracking metrics like resolution rate and token usage. Terminal-Bench is developed by harbor-framework and is written in Python. It is distributed under the Apache 2.0 license, making it freely available for use and modification. The project aims to track progress at the frontier of agent development by offering a standardized way to assess and compare agent capabilities across terminal-based workflows.
The benchmark includes detailed task data and performance metrics accessible through its leaderboard interface. Researchers and developers can run their own agents against the benchmark suite to evaluate their implementations. The tool provides infrastructure for both benchmarking existing models and contributing new tasks to the benchmark suite, supporting the continued evolution of agent evaluation standards.
Checked 11 November 2026.
Nobody on Stackness runs Terminal-Bench yet.
Sign up and add itPosts about Terminal-Bench
- Sonnet 5.5 vs Opus 5.5: which slot each one takes in a coding agent stack
Sonnet 5.5 costs half of Opus 5.5 per token, but cache reads cost $0.20 on both. Which model plans, which one edits, and what one metered run saves.
- Coding agent harness, and which parts of it actually change the output
A harness is everything in a coding agent except the model. On strong models, two September 2026 studies found it moves the bill more than the pass rate.