Skip to content

Benchmark

Terminal-Bench 4.0

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

Models and systems measured
44
Scores recorded
136
Sources
2
IQ weight
1.66

Score distribution

40%
Mean
41.8%
Median
16.8%
Std dev
14.5%
10th percentile
57.9%
90th percentile
1%
Min
1%70.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.66 for chat models, measured over 35 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model or systemScore
51GPT-6.1 Terra· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim50.3%
52GPT-6 AstraCode · low effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.7%
53Claude Fable 5.1Non-code · high effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.4%
54GPT-6 Soladaptive thinking · max effort · default sampling · 5-trial mean · context ≤1M unless noted · public leaderboard · Codex CLI · max· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.4%
55GPT-6 SolTerminal-Bench 4.0 official leaderboard · Codex · max effort · 95% CI ±3.2 · 5.0B tokens · $1.5k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs49.4%
56ERNIE 5.5 Flash Preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim46.5%
57Claude Opus 5Terminal-Bench 4.0 official leaderboard · Claude Code · medium effort · 95% CI ±3.8 · 3.7B tokens · $3.2k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs44.8%
58Claude Fable 5· LLM Stats data44.5%
59Claude Fable 5Non-code · max effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.44.5%
60Claude Fable 5OpenAI launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.44.5%
61Claude Fable 5Terminal-Bench 4.0 official leaderboard · Claude Code · max effort · 95% CI ±3.8 · 3.8B tokens · $7.3k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs44.5%
62Claude Fable 5.1Non-code · medium effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.43.4%
63Claude Fable 5.1Terminal-Bench 4.0 official leaderboard · Claude Code · low effort · 95% CI ±3.6 · 1.3B tokens · $2.4k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs43.3%
64DeepSeek V4.1 Flash 1017· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim42.3%
65DeepSeek V4.1 Flash 1017DeepSeek 1017 first-party evaluation · reasoning effort, candidate sampling, exact run count, and unspecified harnesses not published · Claude Code harness · provider run, distinct from maintainer leaderboard· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.42.3%
66Claude Fable 5Anthropic Fable/Mythos 5.1 launch table· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.42%
67Claude Fable 5System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials · context ≤1M unless noted · Fable· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.42%
68GLM 5.3· LLM Stats data41.8%
69GLM 5.3Terminal-Bench 4.0 official leaderboard · Claude Code · max effort · 95% CI ±3.2 · 8.7B tokens · $2.7k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs41.8%
70Mistral Large 4· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim41%
71Claude Fable 5.1Non-code · low effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.40.2%
72GPT-5.6 SolDeepSeek Harness / max effort· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.39.9%
73GPT-5.6 SolDeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.39.9%
74Claude Haiku 5.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim39.2%
75Claude Haiku 5.5Claude Code --bare · max · 10 trials/task · no internet · safeguards stop trial· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.39.2%
76LongCat 2.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim38.5%
77MiniMax M3.1 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim38.1%
78MiniMax M3.1 FlashMiniMax M3.1 Flash official-weights model-card run · reasoning effort, sampling, and run count not stated for this benchmark · Codex 0.149.1 · provider-reported result, distinct from maintainer leaderboard· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.38.1%
79Grok 4.7· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim38%
80Grok 4.7Harbor · provider harness · task success rate · xhigh effort· Reported by SpaceXAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.38%
81GLM 5.3DeepSeek Harness / max effort· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.9%
82GLM 5.3DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.9%
83Grok 4.7Grok Build · release-page value· Reported by SpaceXAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.6%
84Grok 4.7Terminal-Bench 4.0 official leaderboard · Grok Build · xhigh effort · 95% CI ±3.5 · 5.5B tokens · $3.68k (cost partial: 324/330 trials)· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs37.6%
85GPT-5.6 Sol· LLM Stats data37.3%
86GPT-5.6 SolAnthropic Fable/Mythos 5.1 launch table· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.3%
87GPT-5.6 SolGoogle Model Card· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.3%
88GPT-5.6 SolHarbor · provider harness · task success rate · max effort· Reported by SpaceXAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.3%
89GPT-5.6 SolNon-code · max effort · chart displays 37.3% (raw 37.2727%)· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.3%
90GPT-5.6 SolOpenAI launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37.3%
91GPT-5.6 SolTerminal-Bench 4.0 official leaderboard · Codex · max effort · 95% CI ±3.8 · 4.4B tokens · $2.5k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs37.3%
92GPT-5.6 SolSystem Card capability summary · adaptive thinking · max effort · default sampling · 5 trials · context ≤1M unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.37%
93GLM 5.3 FlashTerminal-Bench 4.0 official leaderboard · Claude Code · none effort · 95% CI ±3.5 · 2.9B tokens · $118.74· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs35.8%
94MiMo V2.6 Pro· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim34.9%
95Claude Opus 5Terminal-Bench 4.0 official leaderboard · Claude Code · low effort · 95% CI ±3.9 · 2.7B tokens · $2.4k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs34.8%
96DeepSeek V4.1 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim31.2%
97DeepSeek V4.1 FlashDeepSeek Harness / max effort· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.31.2%
98DeepSeek V4.1 FlashDeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.31.2%
99MiMo V2.6 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim28.8%
100GPT-5.6 SolNon-code · x-high effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.28.5%

Showing 51–100 of 133 results.

Not ranked

The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.

Model or systemScore
Grok 4.6Harbor · provider harness · task success rate · high effort· Reported by SpaceXAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.20.3%
Grok 4.6Terminal-Bench 4.0 official leaderboard · Grok Build · high effort · 95% CI ±3.1 · 4.0B tokens · $3.6k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs20.3%
Qwen3.8 Max 0902Terminal-Bench 4.0 official leaderboard · Claude Code · max effort · 95% CI ±3.8 · 3.0B tokens · $4.6k· Reported by Terminal-Bench / Stanford / Harbor / Laude Institute· via BenchBoard docs27%

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. · Data by LLM Stats