Terminal-Bench 2 leaderboard
As of November 8, 2026, DeepSeek V4.1 Flash (DeepSeek) leads Terminal-Bench 2 at 90.6%, ahead of MiMo-V2.6-Pro at 89.9% and Gemini 3.8 Flash at 89.4%. 18 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.
What is Terminal-Bench 2? Terminal-Bench 2 tests AI agents on hard, realistic jobs done entirely in a command-line terminal: building software, setting up machine learning, security work, data processing. Made by The Terminal-Bench team (Stanford University and the Laude Institute), November 2025.
Top models on Terminal-Bench 2, November 2026
The May 2026 revision that repairs 28 of the 89 tasks.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash DeepSeek | pass@1 | $0.3 / $1.2 | 1M | Vendor-reported DeepSeek |
| 2 | MiMo-V2.6-Pro Xiaomi MiMo | as reported | $0.435 / $0.87 | 1M | Vendor-reported Xiaomi |
| 3 | Gemini 3.8 Flash Google DeepMind | as reported | $0.75 / $3.75 | 1M | Vendor-reported Google DeepMind |
| 4 | Kimi K3 Moonshot (Kimi) | max effort | $3 / $15 | 1M | Vendor-reported Moonshot AI |
| 5 | GLM-5.3 Zhipu (Z.ai / GLM) | as reported | $1.4 / $4.4 | 1M | Vendor-reported Z.ai (Zhipu) |
| 6 | DeepSeek V4 Pro DeepSeek | pass@1 | $1.32 / $3.96 | 1M | Vendor-reported DeepSeek |
| 7 | MiMo-V2.6-Flash Xiaomi MiMo | as reported | $0.14 / $0.28 | 1M | Vendor-reported Xiaomi |
| 8 | Qwen3.8 Max Alibaba (Qwen) | as reported | $2 / $6 | 1M | Vendor-reported Alibaba (Qwen) |
| 9 | Gemini 3.7 Flash Deprecated Google DeepMind | as reported | $0.75 / $3.75 | 1M | Vendor-reported Google DeepMind |
| 10 | DeepSeek V4 Flash Deprecated DeepSeek | pass@1 | $0.44 / $1.32 | 1M | Vendor-reported DeepSeek |
| 11 | GLM-5.2 Zhipu (Z.ai / GLM) | as reported | $1.4 / $4.4 | 1M | Vendor-reported Z.ai (Zhipu) |
| 12 | Gemini 3.6 Flash Google | as reported | $0.75 / $3.75 | 1M | Vendor-reported Google DeepMind |
| 13 | Gemini 3.5 Flash Deprecated Google | Terminus-2 harness | $1.5 / $9 | 1M | Vendor-reported Google DeepMind |
| 14 | Qwen3.7 Max Alibaba (Qwen) | as reported | $2.5 / $7.5 | 1M | Vendor-reported Alibaba (Qwen) |
| 15 | Gemini 3.1 Pro Preview Preview Google · listed as “Gemini 3.1 Pro” | Terminus-2 harness | $2 / $12 | 1M | Vendor-reported Google DeepMind |
| 16 | Gemini 3.5 Flash-Lite Google | Terminus-2 harness | $0.3 / $2.5 | 1M | Vendor-reported Google DeepMind |
| 17 | Gemini 3.1 Flash-Lite Deprecated Google | Terminus-2 harness | $0.25 / $1.5 | 1M | Vendor-reported Google DeepMind |
Terminal-Bench 2 under other settings
Labs report Terminal-Bench 2 under different settings, and a score from one setting is not comparable with a score from another. Each setting has its own table, ranked on its own.
Terminal-Bench 2.0
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro Preview Preview Google · listed as “Gemini 3.1 Pro” | Terminus-2 harness, thinking high | $2 / $12 | 1M | Vendor-reported Google DeepMind |
| 2 | Gemini 3 Pro Preview Retired Google · listed as “Gemini 3 Pro” | Terminus-2 harness, thinking high | $2 / $12 | 1M | Vendor-reported Google DeepMind |
Terminal-Bench 2: score against price
Can you trust Terminal-Bench 2 scores?
No independent leaderboard publishes Terminal-Bench 2 results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.
- Version 2.1 (May 2026) repairs 28 of the 89 tasks, so 2.0 and 2.1 scores are not comparable. They are separate tables here.
- The agent harness matters: the authors show the same model scoring differently under different agents (Terminus 2, Claude Code, Codex CLI, Gemini CLI).
- Tasks could pull outside dependencies over the internet, and some broke as those drifted; that is one reason for 2.1.
How Terminal-Bench 2 works
The task
Each task drops the agent into a container modelled on a real workflow. It works only through the terminal until the environment reaches the required end state.
The score
Task-specific tests check the final state of the container; a task passes or fails, and the score is the share of tasks passed.
Test set
89 tasks, as published by the authors.
Made by
The Terminal-Bench team (Stanford University and the Laude Institute), first published November 2025.
Official sources
- www.tbench.ai
- Paper: arxiv.org/abs/2601.11868
- Definition read on 2026-11-08
Sources and freshness
19 published scores from 10 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.
- Vendor-reported DeepSeek
- Vendor-reported Alibaba (Qwen)
- Vendor-reported Google DeepMind
- Vendor-reported Google DeepMind
- Vendor-reported Google DeepMind
- Vendor-reported Google DeepMind
- Vendor-reported Xiaomi
- Vendor-reported Z.ai (Zhipu)
- Vendor-reported Google DeepMind
- Vendor-reported Moonshot AI
Terminal-Bench 2 FAQ
Answers derived from the 19 published scores on this page and the current price index.