Skip to content

Benchmark

Terminal-Bench 2.0 · Terminus-2

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.

Models measured
25
Scores recorded
44
Sources
1
IQ weight
No weight published

Score distribution

58.1%
Mean
57.8%
Median
10.3%
Std dev
46.4%
10th percentile
70.3%
90th percentile
37.1%
Min
37.1%82.7%
The band covers the middle 80% of scores. The line marks the median.

Leaderboard

#ModelScore
1GPT 5.5xhigh· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.82.7%
2GPT-5.3 CodexOpenAI GPT-5.4 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.77.3%
3GPT-5.3 Codexxhigh· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.77.3%
4GPT 5.4OpenAI GPT-5.4 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.75.1%
5Qwen3.7 PlusQwen3.7-Plus release · Harbor/Terminus-2 · 5h · 12 CPU/24 GB · temp 1.0 · top_p 0.95 · top_k 20 · max_tokens 80K · 256K · avg@5· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.70.3%
6Qwen3.7 MaxQwen3.7-Max release · Harbor/Terminus-2 · 5h · 12 CPU/24 GB · temp 1.0 · top_p 0.95 · top_k 20 · max_tokens 80K · 256K · avg@5· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.69.7%
7Claude Opus 4.7Harbor · Terminus-2 · thinking disabled · 89 tasks · avg@5· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.69.4%
8GLM 5.1GLM-5.1 official release table · best self-reported · Claude Code· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.69%
9Gemini 3.1 Pro PreviewTerminus-2· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.68.5%
10Kimi K2.6Kimi K2.6 official benchmark table · thinking · temp 1.0 · top_p 1.0 · 262,144 context · Terminus-2 · JSON parser · preserve thinking · avg@10· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.66.7%
11Claude Opus 4.6Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.4%
12GPT 5.2Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.64.7%
13GLM 5.1GLM-5.1 official release table · Terminus-2· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.63.5%
14GPT 5.2OpenAI GPT-5.4 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.62.2%
15Qwen3.6 PlusQwen official release table · thinking mode unless benchmark requires otherwise · Harbor/Terminus-2 · 3h · 32 CPU/48GB · 256K · five-run mean· Reported by Alibaba Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.61.6%
16GLM 5GLM-5 official release table · Claude Code 2.1.14 · Z.ai verified set · five-run mean· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.61.1%
17GLM 5GLM-5 official release table · Terminus-2 · Z.ai verified set· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.60.7%
18Claude Opus 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.59.8%
19Qwen3.6 27BQwen official release table · thinking mode unless benchmark requires otherwise · Harbor/Terminus-2 · 3h · 32 CPU/48GB · 256K · five-run mean· Reported by Alibaba Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.59.3%
20Claude Opus 4.5Harbor · Terminus-2 · 128K thinking budget · 1,335 trials· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.59.3%
21Claude Sonnet 4.6Terminus-2 · max effort · no thinking · 5 runs· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.59.1%
22Claude Opus 4.5Harbor · Terminus-2 · 64K thinking budget · 2,225 trials· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.57.8%
23DeepSeek V4 FlashDeepSeek V4 official model card · Think Max· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.56.9%
24DeepSeek V4 FlashDeepSeek V4 official model card · Think High· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.56.6%
25Gemini 3 Pro PreviewOpus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.56.2%
26GLM 5GLM-5 official release table · Claude Code 2.1.14 · original set · five-run mean· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.56.2%
27GLM 5GLM-5 official release table · Terminus-2 · original set· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.56.2%
28Gemini 3 Pro PreviewGemini 3 Pro Model Card · results as of November 2025 · Terminus-2· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54.2%
29Qwen3.6 FlashQwen3.6-35B-A3B official model card · hosted API name Qwen3.6-Flash· Reported by Alibaba Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.51.5%
30Claude Sonnet 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.51%
31Step 3.5 FlashStep 3.5 Flash official headline table · vanilla· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.51%
32Step 3.5 FlashStep 3.5 technical report · Table 16 · baseline · avg@8· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.51%
33Kimi K2.5· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.50.8%
34Step 3.5 FlashStep 3.5 technical report · Table 16 · 100 rounds · avg@8· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.50.4%
35Claude Sonnet 4.5launch-page setting· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.50%
36Step 3.5 FlashStep 3.5 technical report · Table 16 · 2-hour timeout · avg@8· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.7%
37DeepSeek V4 FlashDeepSeek V4 official model card · Non-Think· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.1%
38Step 3.5 FlashStep 3.5 technical report · Table 16 · 16K output · avg@8· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.48%
39Gemini 3 FlashTerminus-2 harness · public leaderboard· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.47.6%
40DeepSeek V3.2Thinking mode · temperature 1.0 · 128K context · Claude Code framework· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.46.4%
41Step 3.5 FlashStep 3.5 technical report · Table 16 · 16K output without pruning · avg@8· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.45.2%
42Claude Haiku 4.5launch-page setting· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.41%
43DeepSeek V3.2Non-thinking mode · temperature 1.0 · 128K context · Terminus framework· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.39.3%
44DeepSeek V3.2Non-thinking mode · temperature 1.0 · 128K context · Claude Code framework· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.1%

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.