Terminal-Bench 2.1
Version 2.1 of Terminal-Bench, which tests agents on realistic command-line tasks inside sandboxed environments.
Actively evaluated by Artificial Analysis on current frontier models; no saturation evidence recorded.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Qwen3.8 2.4T A95B | 82% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Scout | 3.7% | 2026-11-15 | independent | Source ↗ |
| Qwen3.7 Plus | 61% | 2026-11-15 | independent | Source ↗ |
| Phi-4-mini | 0.4% | 2026-11-15 | independent | Source ↗ |
| NVIDIA Nemotron 3 Nano 4B | 3.7% | 2026-11-15 | independent | Source ↗ |
| K2 Horizon 0.9B | 1.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8-Flash-Next | 86.1% | 2026-11-15 | independent | Source ↗ |
| Nemotron Cascade 2 30B A3B | 20.6% | 2026-11-15 | independent | Source ↗ |
| Llama 3.3 Instruct 70B | 4.9% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 8B | 4.1% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Maverick | 7.9% | 2026-11-15 | independent | Source ↗ |
| Granite 4.2 8B | 18.4% | 2026-11-15 | independent | Source ↗ |
| Kimi K2.7 Code | 67.4% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Nano Omni 30B A3B Reasoning | 6.7% | 2026-11-15 | independent | Source ↗ |
| Magistral Medium 1.2 | 12.4% | 2026-11-15 | independent | Source ↗ |
| DeepSeek V4-Pro | 78.7% | 2026-11-15 | independent | Source ↗ |
| A.X-K2 | 39% | 2026-11-15 | independent | Source ↗ |
| Celeris-1 | 11.2% | 2026-11-15 | independent | Source ↗ |
| Ling 3.0 Tiny | 27.7% | 2026-11-15 | independent | Source ↗ |
| Solar Open2 250B | 44.2% | 2026-11-15 | independent | Source ↗ |
| Nanbeige4.1-3B | 1.1% | 2026-11-15 | independent | Source ↗ |
| DeepSeek V4 Flash Vision | 74.2% | 2026-11-15 | independent | Source ↗ |
| gpt-oss-120b | 26.2% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Coder Next | 38.2% | 2026-11-15 | independent | Source ↗ |
| KAT Coder Pro V2 | 70% | 2026-11-15 | independent | Source ↗ |
| Quasar 438B | 69.3% | 2026-11-15 | independent | Source ↗ |
| Agnes 2.5 Pro Beta | 69.7% | 2026-11-15 | independent | Source ↗ |
| Ring-2.6-1T | 43.1% | 2026-11-15 | independent | Source ↗ |
| Inkling 2 Small | 81.3% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-2.6B | 4.5% | 2026-11-15 | independent | Source ↗ |
| Mistral Medium 3.5 | 50.6% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 31B | 43.4% | 2026-11-15 | independent | Source ↗ |
| GLM 5.3 | 83.9% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.8 Flash | 87.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Next 80B A3B (Reasoning) | 6.7% | 2026-11-15 | independent | Source ↗ |
| GPT-5.5 Instant | 34.8% | 2026-11-15 | independent | Source ↗ |
| DiffusionGemma 26B A4B | 12.4% | 2026-11-15 | independent | Source ↗ |
| G9v3-39A5B | 32.6% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.1 Pro | 73.8% | 2026-11-15 | independent | Source ↗ |
| K2 Horizon 375B A23B | 71.9% | 2026-11-15 | independent | Source ↗ |
| Trinity Large Thinking | 20.6% | 2026-11-15 | independent | Source ↗ |
| Hy3 | 64.4% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8 Max (0902) | 88.8% | 2026-11-15 | independent | Source ↗ |
| Motif 3 | 74.9% | 2026-11-15 | independent | Source ↗ |
| Solar Pro 4 | 57.3% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 4B | 21.3% | 2026-11-15 | independent | Source ↗ |
| G9v3-3B | 6% | 2026-11-15 | independent | Source ↗ |
| Command A+ | 22.8% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3.5 Lightning | 24.3% | 2026-11-15 | independent | Source ↗ |
| MiniMax M3 | 65.2% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-9B | 21.3% | 2026-11-15 | independent | Source ↗ |
| GPT-6 Astra | 88.4% | 2026-11-15 | independent | Source ↗ |
| North Mini Code | 35.6% | 2026-11-15 | independent | Source ↗ |
| GLM 5.3 Flash | 84.3% | 2026-11-15 | independent | Source ↗ |
| K-EXAONE 2.0 | 40.4% | 2026-11-15 | independent | Source ↗ |
| K2 Horizon 3.7B | 28.1% | 2026-11-15 | independent | Source ↗ |
| Mistral Small 4 | 21% | 2026-11-15 | independent | Source ↗ |
| K2 Horizon 7B | 44.2% | 2026-11-15 | independent | Source ↗ |
| Granite 4.2 30B | 26.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 397B A17B | 51.3% | 2026-11-15 | independent | Source ↗ |
| Inkling | 55.1% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Pro | 29.6% | 2026-11-15 | independent | Source ↗ |
| Granite 4.2 3B | 13.9% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 2B | 0% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.5 Flash-Lite | 53.6% | 2026-11-15 | independent | Source ↗ |
| Apodex 1.1 | 69.7% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Ultra | 53.9% | 2026-11-15 | independent | Source ↗ |
| Inkling Small | 55.1% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 14B | 9.7% | 2026-11-15 | independent | Source ↗ |
| Magistral Small 1.2 | 4.5% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 26B A4B | 39% | 2026-11-15 | independent | Source ↗ |
| Ling-3.0-flash-Fin | 62.5% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 3B | 0% | 2026-11-15 | independent | Source ↗ |
| Qwen3.6 35B A3B | 41.6% | 2026-11-15 | independent | Source ↗ |
| MiniCPM5-2B | 8.6% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Lite | 16.1% | 2026-11-15 | independent | Source ↗ |
| LongCat 2.0 | 50.2% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-122B-A10B | 47.2% | 2026-11-15 | independent | Source ↗ |
| K2 Think V2 | 15% | 2026-11-15 | independent | Source ↗ |
| Muse Glimmer | 51.7% | 2026-11-15 | independent | Source ↗ |
| K2 Horizon MoVA 36B A4B | 56.6% | 2026-11-15 | independent | Source ↗ |
| MiniCPM-V 4.6 1.3B | 0% | 2026-11-15 | independent | Source ↗ |
| Ling-3.0-flash-VL | 64.4% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 0.8B | 0.4% | 2026-11-15 | independent | Source ↗ |
| Nex-N2-Pro | 67.8% | 2026-11-15 | independent | Source ↗ |
| Muse Spark Open | 76.4% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8 27B | 79.8% | 2026-11-15 | independent | Source ↗ |
| HyperNova 60B 2605 | 18.4% | 2026-11-15 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.No peer-reviewed human baseline has been published for this benchmark.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.task success (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks terminal-bench-2-1 --batch_size autoopencompass --datasets terminal-bench-2-1 --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for Terminal-Bench 2.1, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace