Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Terminal-Bench

Terminal-Bench is a benchmark for AI agents that work in a command-line terminal, where each task gives the agent a real environment and is scored by tests that check whether the job was actually done.

Also called: Terminal-Bench 2.0, tbench

By Divya Sharma · Checked Nov 15, 2026 · How we review

Coding agents increasingly work through a shell: installing packages, editing files, running builds and debugging. Terminal-Bench measures that. It is run by a team hosted by Stanford and the Laude Institute, and its tasks are drawn from real workflows rather than puzzle-style questions.

How it works

Each task comes with its own environment, a human-written reference solution, and tests that verify the result. The agent is dropped into the environment with instructions and must finish the job using the terminal. The score is the share of tasks it resolves. The Terminal-Bench 2.0 paper describes 89 hard tasks and reports that frontier models and agents scored under 65%. The project has since moved on to newer versions, with Terminal-Bench 4.1 on its site at the time of writing.

Model and harness together

The leaderboard lists both a model and an agent for each result, because the same model can score differently in different agent harnesses. When a model maker quotes a Terminal-Bench number, check which harness and version they used. The site also marks its data with a canary string asking that it never appear in training data, to limit benchmark contamination.

How it differs from SWE-bench

SWE-bench asks an agent to fix real issues in open-source repositories. Terminal-Bench covers a wider spread of command-line work, from system setup to data tasks. Both are useful when choosing a coding agent; neither replaces trying the agent on your own projects in a safe sandbox.

Related terms

SWE-bench
SWE-bench is a coding benchmark built from real GitHub issues in popular Python projects; a model or agent must edit the repository so that the issue is fixed and the project's hidden tests pass.
Coding agent
A coding agent is an AI agent that works on software directly: it reads a codebase, edits files, runs commands and tests, and repeats until a change works, rather than only suggesting code for a person to paste.
Agent harness
An agent harness is the software around a model that turns it into a working agent: the loop, tools, system prompt, permissions, context management and any files used to carry progress between sessions.
AI benchmark
An AI benchmark is a fixed, public set of tasks with an agreed scoring method, used to compare models on the same test, for example coding issues solved, exam questions answered or retrieval quality across many datasets.
Agent sandbox
An agent sandbox is an isolated environment, such as a container, virtual machine or operating-system restriction, in which an AI agent runs code and commands so it cannot reach files, networks or data outside set limits.

Sources