benchmarks.site

Public leaderboard · Coding & software engineering

Terminal-Bench

Tests whether AI agents can complete real tasks by typing commands in a computer terminal, like a developer or system administrator would.

Leaderboard

Top 3 right now

  1. 1stOpus 5.5 ( max )65.8% ± 3.1%
  2. 2ndFable 5.2 ( max )65.8% ± 2.8%
  3. 3rdGemini 4 Carbon ( high )64.5% ± 3.2%

What it measures

Hard, realistic tasks an agent must complete in a sandboxed terminal (building, configuring, debugging, data processing), each checked by tests, scored by task success rate per agent and model.

How to read the numbers

Share of Terminal-Bench 4.1 tasks completed (%) with 95% CI — higher is better.

Results

25 rows · higher is better (Resolution rate)CSV
1Opus 5.5 ( max )Claude Code65.8% ± 3.1%$4.7k
1Fable 5.2 ( max )Claude Code65.8% ± 2.8%$16.5k
3Gemini 4 Carbon ( high )Antigravity CLI64.5% ± 3.2%$1.4k
4Sonnet 5.5 ( max )Claude Code63.9% ± 3.0%$7.3k
5GPT-6.1 Sol ( max )Codex60.9% ± 3.0%$0.6k
6GPT-6 Astra ( max )Codex60.3% ± 2.8%$3.3k
7Kimi K3.1 Pro ( max )Claude Code59.7% ± 4.2%$1.1k
8Fable 5.1 ( xhigh )Claude Code57.6% ± 3.3%$4.9k
9Gemini 4 Hydrogen ( high )Antigravity CLI56.4% ± 3.0%$0.8k
10Gemini 4 Argon ( high )Antigravity CLI55.5% ± 2.8%$1.1k
11Opus 5 ( xhigh )Claude Code53.3% ± 3.1%$6.1k
11GPT-6.1 Terra ( max )Codex53.3% ± 3.1%$0.9k
13GPT-6 Sol ( max )Codex50.3% ± 3.2%$1.5k
14Muse Wonder ( xhigh )Muse Code47.0% ± 3.8%$2.5k
15Fable 5 ( max )Claude Code44.8% ± 3.8%$7.3k
16GLM-5.3 ( max )Claude Code41.2% ± 3.3%$2.7k
17Haiku 5.5 ( max )Claude Code40.0% ± 3.6%$1.0k
18Grok 4.7 ( xhigh )Grok Build38.8% ± 3.6%$3.7k
19GPT-5.6 Sol ( max )Codex38.5% ± 3.8%$2.5k
19DeepSeek V4.1 Flash 1017 ( max )Claude Code38.5% ± 3.8%$0.2k
21Hy4 ( max )Claude Code37.9% ± 3.5%$0.3k
22GLM-5.3-Flash ( max )Claude Code37.0% ± 3.6%$0.4k
23Muse Ember ( xhigh )Muse Code30.6% ± 3.8%$0.2k
24Mistral Large 4 ( max )Claude Code27.6% ± 3.9%$0.7k
25Qwen3.8-Max-0902 ( max )Claude Code27.0% ± 3.9%$4.6k

Note: Terminal-Bench 4.1 leaderboard; as-of date not shown on page (latest listed release Nov 12, 2026).

Source: Terminal-Bench (Stanford / Harbor / Laude Institute) · tbench.ai · retrieved 2026-11-17

Markdown · More Coding & software engineering