Terminal-Bench 4.0: Claude Opus 5.5 and Sonnet 5.5 lead terminal operations with 65%

Unlike synthetic evaluations that only validate code responses in isolated boxes, Terminal-Bench puts models in front of environments with real file systems, command line tools (git, bash, compilers, debuggers) and practical execution constraints.

In this article you will see:

• What Terminal-Bench 4.0 requires from agents in Linux environments.
• The methodology with the Mini-SWE-agent harness and approval criteria.
• The official ranking with Claude, Kimi, Gemini and GPT.
• Cost per test metrics and average token consumption.
• How to run and orchestrate agents in terminals in your local workflow.

What Terminal-Bench 4.0 evaluates

The benchmark simulates real tasks faced by software engineers, system administrators and DevOps specialists in the day to day of servers and open source repositories:

• Failure investigation: diagnose why a service failed to start using systemd logs and network utilities.
• Compilation and build: configure toolchains, resolve missing dependencies and fix compilation errors in C, Rust and Python.
• Data and text manipulation: filter and transform large volumes of records using UNIX commands such as grep, sed, awk and jq.
• Repository automation: complex git operations, merge conflict resolution and patch application.

To compare with cybersecurity and reverse engineering tests in the same ecosystem, see also our analysis of the SRE Bench by Vals AI.

How the methodology works with Mini-SWE-agent

All tests are run through the Mini-SWE-agent harness, a controlled environment that grants the agent an interactive bash session with limited permissions inside ephemeral Docker containers.

On each conversation turn, the model:

1. Issues a command in the terminal (e.g.: cat /var/log/nginx/error.log or pytest -q).
2. Receives the stdout and stderr output from the shell.
3. Analyzes the results and plans the next corrective action.
4. Concludes with a finalization command when the task is completed.

The evaluation validates the final state of the operating system by running a deterministic suite of automated tests and assertions that the agent has no access to inspect.

The Terminal-Bench 4.0 Leaderboard

The consolidated results table from Vals AI shows a tight competitive landscape among top model families:

The data shows that Claude Sonnet 5.5 delivers almost the same performance as Opus 5.5, 64.14% versus 65.15%, with a lower financial cost per run. Kimi K3.1 Pro also surprised by placing 3rd at just $4.57 per test, less than half the cost of Claude Opus. For a complementary analysis on benchmarks that test code in real repositories, check out the guide for SWE-bench Pro.

Token Efficiency and Resolution Time

Beyond raw accuracy, Terminal-Bench 4.0 measures the operational behavior of models:

• GPT-6 Astra is the fastest and most token efficient, consuming only 3.01 million tokens and completing the cycle in 35 minutes.
• Claude Sonnet 5.5 uses a longer iterative reasoning process, 58.53 million tokens, actively exploring diagnostic commands before making definitive decisions.
• GPT-6.1 Sol recorded the lowest absolute financial cost, $1.72 per test, although with a resolution rate about 10% lower than the leaders.

How to Run Agents in Parallel Terminals on bridgecode

The main challenge of operating agents in real terminals is preventing one session from locking your environment or creating conflicting changes in your working branch.

With bridgecode, you get dedicated terminals for each CLI, Claude Code, OpenCode, Codex and Antigravity, running in separate git worktrees. This lets you run an agent executing heavy commands in the background while you keep coding without interruptions. Explore the bridgecode ecosystem to put agent squads to work on your projects.