Skip to main content
benchlm.ai

Terminal-Bench Hard

A display-only Artificial Analysis coding metric for agentic coding and terminal use on a harder Terminal-Bench slice.

Benchmark score on Terminal-Bench Hard — June 13, 2026

BenchLM mirrors the published score view for Terminal-Bench Hard. GPT-5.5 leads the public snapshot at 60.6% , followed by Claude Opus 4.8 (58.3%) and GPT-5.4 (57.6%). BenchLM does not use these results to rank models overall.

1Closed

GPT-5.5

OpenAI

60.6%
Overall 89Context 1M
2Closed

Claude Opus 4.8

Anthropic

58.3%
Overall 93Context 1M
3Closed

GPT-5.4

OpenAI

57.6%
Overall 88Context 1.05M
123 modelsCodingCurrentDisplay onlyUpdated June 13, 2026

The published Terminal-Bench Hard snapshot is tightly clustered at the top: GPT-5.5 sits at 60.6%, while the third row is only 3.0 points behind. The broader top-10 spread is 12.1 points, so the benchmark still separates strong models even when the leaders cluster.

123 models have been evaluated on Terminal-Bench Hard. The benchmark falls in the Coding category. This category carries a 20% weight in BenchLM.ai's overall scoring system. Terminal-Bench Hard is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About Terminal-Bench Hard

Year

2026

Tasks

Agentic coding and terminal tasks

Format

Task success rate

Difficulty

Professional software engineering

BenchLM stores Terminal-Bench Hard separately from Terminal-Bench 2.0 because OpenRouter and Artificial Analysis publish it as a distinct benchmark card.

Artificial Analysis model benchmarks

BenchLM freshness & provenance

Version

Terminal-Bench Hard 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Benchmark score table (123 models)

1
GPT-5.5
60.6%
2
Claude Opus 4.8
58.3%
3
GPT-5.4
57.6%
4
Claude Opus 4.7
54.5%
5
Gemini 3.1 Pro
53.8%
6
GPT-5.3 Codex
53.0%
7
GPT-5.4 mini
52.3%
8
Claude Opus 4.7 (Adaptive)
51.5%
9
Qwen3.7 Max
50.8%
10
Claude Opus 4.6
48.5%
11
Qwen3.7 Plus
47.0%
12
GPT-5.2
47.0%
13
Claude Opus 4.5 Thinking
47.0%
14
DeepSeek V4 Pro (Max)
46.2%
15
Claude Sonnet 4.6
46.2%
16
Claude Opus 4.6 (Adaptive)
46.2%
17
GPT-5.1
45.5%
18
Muse Spark
45.5%
19
Kimi K2.6
43.9%
20
Qwen3.6 Plus
43.9%
21
Qwen 3.6 Max (preview)
43.9%
22
GLM-5.1
43.2%
23
GLM-5
43.2%
24
MiMo-V2.5-Pro
43.2%
25
MiniMax M3
42.4%
26
GPT-5.4 nano
42.4%
27
DeepSeek V4 Pro (High)
41.7%
28
Gemini 3 Pro
41.7%
29
Gemini 3.5 Flash
40.9%
30
Qwen3.5 397B (Reasoning)
40.9%
31
Claude Opus 4.5
40.9%
32
MiMo-V2-Pro
40.9%
33
MiniMax M2.7
39.4%
34
DeepSeek V4 Flash (High)
38.6%
35
GPT-5 (medium)
37.9%
36
Grok 4
37.9%
37
Grok 4.3
37.9%
38
GPT-5.2-Codex
37.1%
39
o3
37.1%
40
Nemotron 3 Ultra
36.4%
41
Gemma 4 31B
36.4%
42
DeepSeek V4 Flash (Max)
35.6%
43
Qwen3.5 397B
35.6%
44
Step 3.7 Flash
35.6%
45
Kimi K2.5 (Reasoning)
34.8%
46
GPT-5.1-Codex-Max
34.8%
47
Qwen3.6-27B
34.8%
48
Qwen3.6-35B-A3B
34.8%
49
Kimi K2.5
34.8%
50
MiMo-V2-Omni
34.8%
51
GPT-5.1-Codex
34.8%
52
Claude 4.1 Opus Thinking
34.3%
53
Hy3 Preview
34.1%
54
Mistral Medium 3.5 128B
33.3%
55
GLM-5-Turbo
33.3%
56
GPT-5 (high)
32.6%
57
Qwen3.5-27B
32.6%
58
DeepSeek V3.2
32.6%
59
GLM-5V-Turbo
32.6%
60
GLM-4.7
31.8%
61
Gemini 3 Flash
31.8%
62
Qwen3.5-122B-A10B
31.1%
63
GLM-4.6
28.8%
64
Claude 4 Sonnet
27.3%
65
Gemini 2.5 Pro
26.5%
66
Qwen3.5-35B-A3B
26.5%
67
MiMo-V2-Flash
25.8%
68
DeepSeek V3.1 (Reasoning)
25.0%
69
Command A+
25.0%
70
Gemini 3.1 Flash-Lite
24.2%
71
DeepSeek V3.1
24.2%
72
Grok 4.1 Fast (Reasoning)
24.2%
73
GPT-OSS 120B
23.5%
74
K-Exaone
22.7%
75
Trinity-Large-Preview
22.7%
76
Trinity-Large-Thinking
22.7%
77
Ling 2.6 Flash
21.2%
78
GLM-4.5-Air
20.5%
79
Qwen3 Max
20.5%
80
Grok 4 Fast (Reasoning)
18.9%
81
Gemma 4 12B
18.2%
82
Grok Code Fast 1
17.4%
83
Mistral Small 4 (Reasoning)
17.4%
84
Mistral Small 4
17.4%
85
Mistral Large 3
15.9%
86
Kimi K2
15.9%
87
DeepSeek-R1
15.9%
88
Grok 4.1 Fast
14.4%
89
GPT-4.1
13.6%
90
Gemma 4 26B A4B
13.6%
91
o1
12.9%
92
Gemini 2.5 Flash
12.1%
93
Nemotron 3 Nano 30B
12.1%
94
GPT-OSS 20B
10.6%
95
GPT-4o
8.3%
96
Nemotron 3 Nano Omni 30B A3B
8.3%
97
Gemma 4 E4B
8.3%
98
GPT-4.1 mini
7.6%
99
o3-mini
6.8%
100
Llama 3.1 405B
6.8%
101
DeepSeek V3
6.8%
102
Llama 4 Maverick
6.8%
103
Mistral Large 2
6.1%
104
Nova Pro
6.1%
105
LFM2.5-8B-A1B
4.5%
106
Solar Pro 2
4.5%
107
Phi-4
3.8%
108
GPT-4.1 nano
3.8%
109
Gemma 3 27B
3.8%
110
Mistral Medium 3
3.8%
111
Gemma 4 E2B
3.0%
112
Nemotron Ultra 253B
2.3%
113
Sarvam 30B
2.3%
114
Sarvam 105B
1.5%
115
Llama 4 Scout
1.5%
116
Exaone 4.0 32B
1.5%
117
Claude 3 Haiku
0.8%
118
Granite-4.0-1B
0.0%
119
Granite-4.0-H-1B
0.0%
120
Exaone 4.0 1.2B
0.0%
121
LFM2.5-VL-1.6B-Extract
0.0%
122
Granite-4.0-350M
0.0%
123
Granite-4.0-H-350M
0.0%

FAQ

What does Terminal-Bench Hard measure?

A display-only Artificial Analysis coding metric for agentic coding and terminal use on a harder Terminal-Bench slice.

Which model scores highest on Terminal-Bench Hard?

GPT-5.5 by OpenAI currently leads with a score of 60.6% on Terminal-Bench Hard.

How many models are evaluated on Terminal-Bench Hard?

123 AI models have been evaluated on Terminal-Bench Hard on BenchLM.

Compare Top Models on Terminal-Bench Hard

GPT-5.5 vs Claude Opus 4.8Claude Opus 4.8 vs GPT-5.4GPT-5.4 vs Claude Opus 4.7Claude Opus 4.7 vs Gemini 3.1 Pro
Last updated: June 13, 2026 · BenchLM version Terminal-Bench Hard 2026

The AI models change fast. We track them for you.

For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.