reasoning benchmark
Terminal-Bench
Terminal-Bench is a benchmark for testing AI agents in real terminal environments. It evaluates how well agents can handle real-world, end-to-end tasks autonomously, including compiling code, training models, setting up servers, system administration, security tasks, data science workflows, and cybersecurity vulnerabilities. The benchmark consists of a dataset of ~100 hand-crafted, human-verified tasks and an execution harness that connects language models to a terminal sandbox.
Updated Aug 7, 2026
Published models25
Registry coverage25
MetricScore
EvidenceB
Terminal-Bench leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | AN | 50.0% | 100.0% | 25 | C | |
| 2 | MI | 47.9% | 95.8% | 25 | C | |
| 3 | MA | 47.1% | 91.7% | 25 | C | |
| 4 | MI | 46.3% | 87.5% | 25 | C | |
| 5 | AN | 43.3% | 83.3% | 25 | C | |
| 6 | AM Nova 2 ProAmazon | 41.3% | 79.2% | 25 | C | |
| 7 | AN | 41.0% | 75.0% | 25 | C | |
| 8 | ZA | 40.5% | 70.8% | 25 | C | |
| 9 | ME | 39.5% | 66.7% | 25 | C | |
| 10 | AN | 39.2% | 62.5% | 25 | C | |
| 11 | DE | 37.7% | 58.3% | 25 | C | |
| 12 | ZA | 37.5% | 54.2% | 25 | C | |
| 13 | AN | 35.5% | 50.0% | 25 | C | |
| 14 | AN | 35.2% | 45.8% | 25 | C | |
| 15 | ME | 33.8% | 41.7% | 25 | C | |
| 16 | ZA | 33.3% | 37.5% | 25 | C | |
| 17 | AMNova 2 LiteAmazon | 32.5% | 33.3% | 25 | C | |
| 18 | DE | 31.3% | 29.2% | 25 | C | |
| 19 | XI | 30.5% | 25.0% | 25 | C | |
| 20 | ZA | 30.0% | 20.8% | 25 | C | |
| 21 | MA | 30.0% | 16.7% | 25 | C | |
| 22 | NV | 25.8% | 12.5% | 25 | C | |
| 23 | MA | 25.0% | 8.3% | 25 | C | |
| 24 | NV | 8.5% | 4.2% | 25 | C | |
| 25 | DE | 5.7% | 0.0% | 25 | C |
Score distribution
Top published rows on the benchmark's original scale.
Terminal-Bench highlights
The top published results on this benchmark's own scale.
What is Terminal-Bench?
Definition and scoring fields from the benchmark registry.
Terminal-Bench is a benchmark for testing AI agents in real terminal environments. It evaluates how well agents can handle real-world, end-to-end tasks autonomously, including compiling code, training models, setting up servers, system administration, security tasks, data science workflows, and cybersecurity vulnerabilities. The benchmark consists of a dataset of ~100 hand-crafted, human-verified tasks and an execution harness that connects language models to a terminal sandbox.
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- Terminal-Bench
- Modality
- text
- Primary category
- reasoning
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- terminal-bench|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about Terminal-Bench.
Which model scores highest on Terminal-Bench?
Claude Sonnet 4.5 is currently ranked first with 50.0%.
What does Terminal-Bench measure?
Terminal-Bench is a benchmark for testing AI agents in real terminal environments. It evaluates how well agents can handle real-world, end-to-end tasks autonomously, including compiling code, training models, setting up servers, system administration, security tasks, data science workflows, and cybersecurity vulnerabilities. The benchmark consists of a dataset of ~100 hand-crafted, human-verified tasks and an execution harness that connects language models to a terminal sandbox.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
25 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.