benchgap
Agentic · terminal

Terminal-Bench 4.0 leaderboard

As of 2026-11-18, the highest measured score on Terminal-Bench 4.0 is 70.6% by Claude Sonnet 5.5. 35 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#ModelScoreSource
1Claude Sonnet 5.570.6%measured
2Claude Mythos 5.270.5%measured
3GPT-6.2 Astra67.7%estimated ± 6.1 pp, medium confidence
4Claude Fable 5.267.6%measured
5Gemini 4 Carbon67.5%estimated ± 3.7 pp, high confidence
6Claude Opus 5.566.4%measured
7Kimi K3.1 Pro65.7%measured
8GPT-6.2 Sol63.3%estimated ± 6.1 pp, medium confidence
9GPT-6.1 Sol61.5%estimated ± 4.7 pp, high confidence
10Claude Mythos 5.160.9%measured
11Muse Wonder Full59.5%estimated ± 6.1 pp, medium confidence
12Gemini 4 Hydrogen58.3%estimated ± 3.7 pp, high confidence
13GPT-6 Astra57.9%measured
14Gemini 4 Argon57.4%measured
15Claude Opus 556.0%estimated ± 4.7 pp, high confidence
16Claude Fable 5.155.8%measured
17GPT-6 Sol53.7%estimated ± 4.7 pp, high confidence
18Muse Wonder52.5%estimated ± 3.7 pp, high confidence
19Pareto 26.951.0%measured
20Pareto 26.10 Preview50.8%measured
21GPT-6.1 Terra50.3%measured
22Claude Fable 549.3%estimated ± 4.7 pp, high confidence
23Qwen4 Max48.5%measured
24ERNIE 5.5 Flash Preview46.5%measured
25Ling 3.1 Flash46.0%measured
26GLM-5.345.3%estimated ± 3.7 pp, high confidence
27Hy444.4%measured
28GPT-5.6 Sol43.8%estimated ± 4.7 pp, high confidence
29GLM-5.3-Flash42.5%estimated ± 4.7 pp, high confidence
30DeepSeek V4.1 Flash 101742.3%measured
31Muse Ember40.9%measured
32Claude Haiku 5.539.2%measured
33LongCat-2.538.5%measured
34MiniMax M3.1 Flash38.1%measured
35Grok 4.738.0%measured
36Inkling 2 Small35.7%estimated ± 6.1 pp, low confidence
37Qwen3.8-27B35.7%estimated ± 6.1 pp, low confidence
38MiMo-V2.6-Pro34.9%measured
39Step 5 Preview33.3%measured
40Mistral Large 430.9%estimated ± 3.7 pp, medium confidence
41Claude Opus 4.829.4%estimated ± 4.7 pp, high confidence
42MiMo-V2.6-Flash28.8%measured
43Kimi K328.7%estimated ± 4.7 pp, high confidence
44Ember-127.4%estimated ± 4.7 pp, high confidence
45SWE-227.3%measured
46GPT-5.6 Terra27.1%estimated ± 4.7 pp, high confidence
47Grok 4.625.1%estimated ± 4.7 pp, high confidence
48GPT-5.6 Luna22.8%estimated ± 4.7 pp, high confidence
49SWE-2 Flash20.0%measured
50Muse Spark 1.319.6%estimated ± 4.7 pp, medium confidence
51GPT-6 Luna19.2%estimated ± 3.7 pp, medium confidence
52Hy4 flash19.2%measured
53Gemini 3.8 Flash19.1%measured
54Gemini 3.7 Flash16.6%estimated ± 4.7 pp, medium confidence
55Grok 4.516.6%estimated ± 4.7 pp, medium confidence
56Claude Sonnet 516.2%estimated ± 4.7 pp, medium confidence
57Beam 1.0 501B-A23B10.6%estimated ± 3.7 pp, medium confidence
58MiniMax M310.5%estimated ± 4.7 pp, medium confidence
59DeepSeek V4 Pro9.3%estimated ± 4.7 pp, medium confidence
60Gemini 3.1 Pro3.9%estimated ± 4.7 pp, medium confidence
61GPT-5.4 mini2.6%estimated ± 4.7 pp, medium confidence
62Inkling-Small1.3%estimated ± 4.7 pp, medium confidence
63Nemotron 3.5 Super VL 120B A12B0.4%estimated ± 4.7 pp, medium confidence
64Inkling0.0%estimated ± 4.7 pp, medium confidence

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General