Mini toolsTerminal-Bench 2.1
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
Terminal-Bench 2.1 evaluates AI agents on 89 hard, realistic tasks in command-line terminal environments inspired by real workflows, spanning domains such as compiling code, training models, and setting up servers. Each task features a unique environment, a human-written reference solution, and comprehensive automated tests for verification; frontier models currently score under 65%.
Agentic
Text
Accuracy
Max 100.0%
Released Jan 2026
5
Results
5
Models scored
88.8%
Top: GPT-5.6 Sol
82.5%
Median
Best results
Top primary scores; one row per model.
Frontier over time
Each dot is one model result; the line traces the running best score.
All results
Showing all configurations including non-primary alternates. · Show only primary
| # | Model | Score | Conditions | Eval date | Source | Flags |
|---|---|---|---|---|---|---|
| 1 | GPT-5.6 Sol | 88.8% | agentic | 26 Jun 2026 | Self-reported | Primary |
| 2 | GPT 5.6 Terra | 84.3% | agentic | 26 Jun 2026 | Self-reported | Primary |
| 3 | GLM 5.2 | 82.7% | 0-shot · agentic | — | Self-reported | |
| 4 | GPT 5.6 Luna | 82.5% | agentic | 26 Jun 2026 | Self-reported | Primary |
| 5 | GLM 5.2 | 81.0% | 0-shot · standard | — | Self-reported | Primary |
| 6 | Claude Sonnet 5 | 80.4% | agentic | 30 Jun 2026 | Self-reported | Primary |
