BenchAtlas

Benchmarks / Agentic Coding

Terminal-Bench

agent + model

End-to-end terminal tasks executed by agent+model pairs.

maintained by Laude Institute / Harbor · www.tbench.ai · 496 observations on record

Agent harness required — results on this benchmark measure whole agent+model systems (scaffold, tools, budgets), not bare models. Scores here are never merged with bare-model numbers, and only standardized canonical harnesses can carry a family score into the BenchAtlas Index.

Versions

4 versions
VersionTasks
Terminal-Bench 2.0 current
terminal-bench-2-0
—
Terminal-Bench 1.0
terminal-bench-1-0
—
Terminal-Bench 2.1
terminal-bench-2-1
—
Terminal-Bench Hard
terminal-bench-hard
—

Current leaderboard

Terminal-Bench 2.0 · Accuracy · top 25 of 132 systems
7377808487
nexau + GPT 5.5 (2026-04-22)
84.7
capy-build + GPT 5.5 (2026-04-22)
83.2
Codex CLI + GPT 5.5 (2026-04-22)
82.3
claude-wozcode-plugin + Claude Opus 4.7
80.2
judy + Gemini 3.1 Pro Preview
80.2
sage + GPT 5.3 Codex
78.4
Droid + GPT 5.3 Codex
77.3
terminus-kira-env-bootstrap + Claude Opus 4.6
76.4
ruley + GPT 5.3 Codex
75.8
codelia + GPT 5.3 Codex
75.7
#SystemAccuracy %
1agent + model nexau + GPT 5.5 (2026-04-22)
OpenAI
84.7%
2agent + model capy-build + GPT 5.5 (2026-04-22)
OpenAI
83.2%
3agent + model Codex CLI + GPT 5.5 (2026-04-22)
OpenAI
82.3%
4agent + model claude-wozcode-plugin + Claude Opus 4.7
Anthropic
80.2%
5agent + model judy + Gemini 3.1 Pro Preview
Google DeepMind
80.2%
6agent + model sage + GPT 5.3 Codex
OpenAI
78.4%
7agent + model Droid + GPT 5.3 Codex
OpenAI
77.3%
8agent + model terminus-kira-env-bootstrap + Claude Opus 4.6
Anthropic
76.4%
9agent + model ruley + GPT 5.3 Codex
OpenAI
75.8%
10agent + model codelia + GPT 5.3 Codex
OpenAI
75.7%
11agent + model capy-build + Claude Opus 4.6
Anthropic
75.3%
12agent + model simple_codex + GPT 5.3 Codex
OpenAI
75.1%
13agent + model terminus-3-3 + Gemini 3.1 Pro Preview
Google DeepMind
74.8%
14agent + model terminus-3-3 + Claude Opus 4.6
Anthropic
74.7%
15agent + model mux + GPT 5.3 Codex
OpenAI
74.6%
16agent + model final + Claude 4.6 Opus
Anthropic
72.1%
17agent + model judy + Claude Opus 4.6
Anthropic
71.9%
18agent + model spoox-m + GPT 5.3 Codex
OpenAI
71.5%
19agent + model Droid + Claude Opus 4.6
Anthropic
69.9%
20agent + model ante + Gemini 3 Pro Preview
Google DeepMind
69.4%
21agent + model IndusAGICodingAgent + GPT 5.3 Codex
OpenAI
69.1%
22agent + model Crux + Claude Opus 4.6
Anthropic
66.9%
23agent + model mux + Claude Opus 4.6
Anthropic
66.5%
24agent + model deepagent-harbor + GPT 5.2 Codex
OpenAI
66.5%
25agent + model clnkr + GPT 5.5 (2026-04-22)
OpenAI
66.1%

One row per evaluated system — reasoning-effort variants rank separately. Where a system has attr variants for this version (eval splits, style control), only the canonical variant is ranked: split=semi_private, style_control=true, or the unsplit run.

Source agreement

where evaluating organizations agree — and don't
Terminal-Bench 2.1Artificial Analysis vs Terminal-Bench Leaderboard
overlap 5 systemsrank correlation 0.80mean |Δ| 5.27max |Δ| 13.43
most contested: Claude Opus 4.7 (+13.43) · Claude Opus 4.8 (+5.76) · GLM 5.1 (+3.15)

Possible reasons: different agent harnesses ran the tasks; verification differs (independent vs community); run setting 'verified' differs (— vs true); run setting 'integration_method' differs (— vs API).