# Terminal-Bench 4.0 ![Category](https://img.shields.io/badge/Category-Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Version](https://img.shields.io/badge/Version-4.0-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) > **Quick answer:** Terminal-Bench 4.0 is an agentic terminal/coding benchmark evaluating how well AI models complete multi-step tasks in a shell environment. Claude Mythos 5.1 leads at 60.9%, with Claude Fable 5.1 close behind at 55.8%, both ahead of Claude Opus 5 (52.3%) and GPT-5.6 Sol (37.3%), as reported by Anthropic in September 2026. ## At a Glance **What it tests:** An AI agent's ability to complete real, multi-step terminal/coding tasks — reading and modifying code, running shell commands, and verifying task completion — within a sandboxed environment. **Why it matters:** Terminal Bench 4.0 is the version of the Terminal-Bench task suite referenced in Anthropic's September 2026 Claude Fable 5.1 / Claude Mythos 5.1 launch, providing a fresher, currently-unsaturated read on agentic coding/terminal capability than earlier Terminal-Bench releases. **Known limitations:** Anthropic's announcement does not publicly detail the task set, authorship, or scoring methodology for this specific version; treat this page as tracking the scores reported in that release rather than an independently verified benchmark specification. It is a distinct, non-comparable benchmark from [TerminalBench 2.1](/benchmarks/tbench/terminalbench-21) and [Terminal-Bench 3.0](/benchmarks/stanford-laude-institute/terminal-bench-30) despite the similar name. ## What Terminal-Bench 4.0 Measures Terminal-Bench 4.0 measures an agent's ability to autonomously complete goal-oriented tasks through a terminal/shell interface, in the same broad tradition as earlier Terminal-Bench releases — file manipulation, code changes, and multi-step CLI workflows, verified against the expected end state of the environment. Anthropic references it as one of the primary agentic-coding benchmarks in its Claude Fable 5.1 / Claude Mythos 5.1 launch, reporting scores for Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, Claude Opus 5, and GPT-5.6 Sol. Notably, Claude Mythos 5.1 — the trusted-access-only, more-permissive-safeguard sibling of Fable 5.1 — reports the highest score of any evaluated model on this version of the benchmark, ahead of Fable 5.1 itself. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent (terminal / coding) | | Metric | % tasks completed successfully | | Version | 4.0 | | Saturation | Low | ## How Terminal-Bench 4.0 Is Scored Agents attempt terminal-based coding and shell tasks and are scored on the percentage of tasks successfully completed, verified via automated checks against the expected end state of the environment — consistent with the scoring approach used by earlier Terminal-Bench versions. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Claude Mythos 5.1](/models/anthropic/claude-mythos-5-1) | 60.9% | [Anthropic: Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) | 2026-09 | | 2 | [Claude Fable 5.1](/models/anthropic/claude-fable-5-1) | 55.8% | [Anthropic: Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) | 2026-09 | | 3 | [Claude Opus 5](/models/anthropic/claude-opus-5) | 52.3% | [Anthropic: Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) | 2026-09 | | 4 | [Claude Fable 5](/models/anthropic/claude-fable-5) | 42.0% | [Anthropic: Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) | 2026-09 | | 5 | [GPT-5.6 Sol](/models/openai/gpt-5-6-sol) | 37.3% | [Anthropic: Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) | 2026-09 | *Scores sourced from Anthropic's Claude Fable 5.1 / Claude Mythos 5.1 announcement, September 2026.* ## Terminal-Bench 4.0 on Benchgen > No Benchgen results yet — [be the first to run Terminal-Bench 4.0](/benchmarks/anthropic/terminal-bench-40). ## Terminal-Bench 4.0 vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **Terminal-Bench 4.0** | Agentic terminal/coding tasks (2026 revision) | Low | | [Terminal-Bench 3.0](/benchmarks/stanford-laude-institute/terminal-bench-30) | Harder, rolling agentic terminal-use tasks (Stanford/Laude Institute) | Low | | [TerminalBench 2.1](/benchmarks/tbench/terminalbench-21) | Earlier-generation agentic terminal-use tasks | Medium | | [Terminal-Bench-Science 0.1](/benchmarks/anthropic/terminal-bench-science-01) | Agentic scientific-research tasks via terminal use | Low | Terminal-Bench 4.0, Terminal-Bench 3.0, and TerminalBench 2.1 are distinct, non-comparable benchmarks despite similar names — each uses its own task set and scoring scale. ## Run Terminal-Bench 4.0 on Your Model Benchgen lets you run agentic terminal/coding tasks against your own model and harness, tracking task-completion rate over time as a complement to vendor-reported Terminal-Bench 4.0 scores. ## Frequently Asked Questions

What is Terminal-Bench 4.0?

Terminal-Bench 4.0 is an agentic terminal/coding benchmark referenced in Anthropic's September 2026 Claude Fable 5.1 / Claude Mythos 5.1 launch, measuring how reliably a model completes multi-step tasks in a shell environment.

What does a good score look like on Terminal-Bench 4.0?

Claude Mythos 5.1 reports the highest score at 60.9%, followed by Claude Fable 5.1 at 55.8%, Claude Opus 5 at 52.3%, Claude Fable 5 at 42.0%, and GPT-5.6 Sol at 37.3%, all as of September 2026.

How does Terminal-Bench 4.0 differ from Terminal-Bench 3.0 and TerminalBench 2.1?

They are distinct, non-comparable benchmarks despite the similar naming — each version uses its own task set and difficulty scale, so scores should not be compared across versions.

Rank	Model	Score
1	claude-opus-5-5	66.4
2	claude-mythos-5-1	60.9
3	gpt-6-astra	57.9
4	claude-fable-5-1	55.8
5	grok-4-7	37.6
6	deepseek-v4-1-flash	31.2

Terminal-Bench 4.0 — Results

Terminal-Bench 4.0 — Evaluations