Terminal-Bench 4.0
Terminal-Bench 4.0 agentic terminal tasks. DeepSeek-V4.1-Flash evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; Pass@1.
Category
Agentic Coding
Score type
percent
Scale
100%
higher is better
Models tested
8
What is Terminal-Bench 4.0?
Terminal-Bench tests whether a model can operate a computer through a terminal as an agent: it gets a Linux container and has to solve hands-on tasks (installing packages, inspecting files, writing scripts) purely via shell commands within a time limit. The score is the share of tasks completed successfully, averaged over several runs.
What does it test?
Terminal-Bench 4.0 measures how well a model works as an autonomous coding agent. In detail, it checks:
- planning and executing multi-step shell workflows
- installing, configuring and debugging software in a container
- handling timeouts and failing commands
- finishing the task within the given compute and time budget
How to read the score
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.
Frequently asked questions
What is the Terminal-Bench 4.0 benchmark?
Terminal-Bench tests whether a model can operate a computer through a terminal as an agent: it gets a Linux container and has to solve hands-on tasks (installing packages, inspecting files, writing scripts) purely via shell commands within a time limit. The score is the share of tasks completed successfully, averaged over several runs.
What does Terminal-Bench 4.0 test?
Terminal-Bench 4.0 mainly measures how well a model works as an autonomous coding agent. Concretely, it checks: planning and executing multi-step shell workflows; installing, configuring and debugging software in a container; handling timeouts and failing commands.
How is the Terminal-Bench 4.0 score calculated?
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.
Where do I get Terminal-Bench 4.0 data and scores?
The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/terminal-bench-4-0/scores) or pull it directly from the Open LLM Radar API with any HTTP client.
Benchmark data & downloads
Where to get, inspect and reuse the data behind this benchmark.
Model scores 8 models tested
| # | Model | Score | Date | Source | Verified |
|---|---|---|---|---|---|
| 1 |
Kimi K3.1 Pro
Moonshot AI
|
65.7%
|
07.11.2026 | Report | Verified |
| 2 |
DeepSeek-V4.1-Flash-1017
DeepSeek
|
42.3%
|
17.10.2026 | Report | Verified |
| 3 |
Mistral Large 4
Mistral AI
|
41%
|
— | HF model card eval results + Mistral blog: Terminal-Bench 4.0 41% (13pp above preview) | Verified |
| 4 |
MiniMax-M3.1-Flash
MiniMaxAI
|
38.1%
|
19.10.2026 | Report | Verified |
| 5 |
GLM-5.3
Zhipu AI
|
37.9%
|
19.10.2026 | Report | Unverified |
| 6 |
DeepSeek-V4.1-Flash
DeepSeek
|
31.2%
|
10.09.2026 | Report | Verified |
| 7 |
Nanbeige4.3-33B
Nanbeige
|
23.8%
|
24.10.2026 | Report | Verified |
| 8 |
Kimi K3
Moonshot AI
|
12.6%
|
19.10.2026 | Report | Unverified |