Terminal-Bench
Agentic terminal tasks; frontier vendors have largely stopped reporting it
2.0agents and tool usesuperseded
Score in %
Definition
- Version
- 2.0
- Category
- agents and tool use
- Maintainer
- Terminal-Bench team (Stanford, Harbor and Laude Institute), USA
- Status
- superseded
- Home page
- https://www.tbench.ai/leaderboard
Sources
- OpenAI Releases GPT-5.5: Agentic Coding Ceiling Tops 14 Benchmarks rits.shanghai.nyu.edu https://rits.shanghai.nyu.edu/ai/openai-releases-gpt-5-5-agentic-coding-ceiling-tops-14-benchmarks/
- openai.com/index/introducing-gpt-5-3-codex https://openai.com/index/introducing-gpt-5-3-codex
- Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces arxiv.org https://arxiv.org/abs/2601.11868
- Terminal-Bench 2.0 Leaderboard & Scores , October 2026 benchlm.ai https://www.benchlm.ai/benchmarks/terminal-bench-2
- TERMINAL-BENCH tbench.ai https://www.tbench.ai/leaderboard
Scores at a Glance
Score in %
Higher is better for this unit.
1
GPT-5.5OpenAI
Claude Fable 5AnthropicGPT-5.3-CodexOpenAIxhigh
DeepSeek-V4-ProDeepSeek
Composer 2Anysphere (Cursor)DeepSeek-V4-FlashDeepSeek
Qwen3.6-35B-A3BAlibaba (Qwen)Qwen3-Coder-NextAlibaba (Qwen)scaffold not stated
Devstral 2 (123B, 2512)Mistral AIQwen3-Coder-480B-A35B-InstructAlibaba (Qwen)Devstral Small 2 (24B, 2512)Mistral AI
GPT-5.5OpenAI82.7 %
vendor-reported2Claude Fable 5Anthropic82 %
vendor-reported3GPT-5.3-CodexOpenAIxhigh77.3 %
vendor-reported4DeepSeek-V4-ProDeepSeek67.9 %
vendor-reported5Composer 2Anysphere (Cursor)61.7 %
vendor-reported6DeepSeek-V4-FlashDeepSeek56.9 %
vendor-reported7Qwen3.6-35B-A3BAlibaba (Qwen)51.5 %
vendor-reported8Qwen3-Coder-NextAlibaba (Qwen)scaffold not stated36.2 %
vendor-reported9Devstral 2 (123B, 2512)Mistral AI32.6 %
vendor-reported10Qwen3-Coder-480B-A35B-InstructAlibaba (Qwen)23.9 %
vendor-reported11Devstral Small 2 (24B, 2512)Mistral AI22.5 %
vendor-reported0255075100
Where this comes from
Each bar is one model, with its most independent recorded score. Hatched bars are vendor-reported; the word beside each bar says who ran it. Settings and dates differ between rows; the Scores table below has every result.
Scores
| Model | Score | Who ran it | Setting | View or population | Date | Source |
|---|---|---|---|---|---|---|
| GPT-5.5 | 82.7 % | vendor-reported | – | – | – | rits.shanghai.nyu.edu |
| Claude Fable 5 | 82 % | vendor-reported | – | – | – | ecosistemastartup.com |
| GPT-5.5 | 82 % | vendor-reported | – | rank 1, BenchLM aggregator, top score | 10 Oct 2026 | benchlm.ai |
| GPT-5.3-Codex | 77.3 % | vendor-reported | xhigh | – | 5 Feb 2026 | openai.com |
| DeepSeek-V4-Pro | 67.9 % | vendor-reported | – | – | – | blog.buildfastwithai.com |
| Composer 2 | 61.7 % | vendor-reported | – | – | – | cursor.com |
| DeepSeek-V4-Flash | 56.9 % | vendor-reported | – | – | – | blog.buildfastwithai.com |
| Qwen3.6-35B-A3B | 51.5 % | vendor-reported | – | – | – | rits.shanghai.nyu.edu |
 MiMo-V2-Flash | 38.5 % | vendor-reported | – | MiMo-V2-Flash model card, post-trained model table | 16 Dec 2025 | huggingface.co |
| Qwen3-Coder-Next | 36.2 % | vendor-reported | scaffold not stated | – | – | huggingface.co |
| Devstral 2 (123B, 2512) | 32.6 % | vendor-reported | – | – | – | huggingface.co |
| Qwen3-Coder-480B-A35B-Instruct | 23.9 % | vendor-reported | – | – | – | huggingface.co |
| Devstral Small 2 (24B, 2512) | 22.5 % | vendor-reported | – | – | – | huggingface.co |
Where this comes from
Scores from different settings, harnesses or dates are not directly comparable; the row says which. Sorted by score, highest first, as a reading aid, not a ranking.
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Terminal-Bench", https://aurion.semai.id/benchmarks/terminal-bench-2-0/, records of October 2026.