Agents · the terminal · reference
Terminal-Bench 4.0
An agent gets a shell and a goal: build software, debug a service, configure a system. Version 4.0 is the current and hardest set.
- Models
- 9
- By makers
- 9
- Latest
- Oct 13
| # | Model | Bar | Score | Source |
|---|---|---|---|---|
| 1 | Claude Sonnet 5.5Anthropic · $4.00/1M · Coming soon | 70.6%vSep 28, 2026 | MakerSep 28, 2026 | |
| 2 | Claude Opus 5.5Anthropic · $8.00/1M · Coming soon | 66.4%vSep 22, 2026 | MakerSep 22, 2026 | |
| 3 | Claude Opus 5Anthropic · $10.00/1M · Coming soon | 52.3%vSep 22, 2026 | MakerSep 22, 2026 | |
| 4 | Ling 3.1 FlashinclusionAI · $0.45/1M · Coming soon | 46%vOct 13, 2026 | MakerOct 13, 2026 | |
| 5 | MiMo V2.6 ProXiaomi · $0.54/1M · Coming soon | 34.9%vSep 21, 2026 | MakerSep 21, 2026 | |
| 6 | Step 5 PreviewStepFun · $1.43/1M · Coming soon | 33.3%vSep 20, 2026 | MakerSep 20, 2026 | |
| 7 | DeepSeek V4.1 FlashDeepSeek · $0.30/1M · Coming soon | 31.2%vSep 10, 2026 | MakerSep 10, 2026 | |
| 8 | MiMo V2.6 FlashXiaomi · $0.17/1M · Coming soon | 28.8%vSep 21, 2026 | MakerSep 21, 2026 | |
| 9 | Claude Sonnet 5Anthropic · $6.00/1M · Coming soon | 10.3%vSep 28, 2026 | MakerSep 28, 2026 |
% of tasks passed. v marks a score the model’s maker reported. Prices are FullsendAI’s per million tokens, blended 3 input : 1 output.
About the data
From Terminal-Bench. Code and tasks are Apache-2.0; tbench.ai publishes no licence for its leaderboard data.