Terminal-Bench
Agentic terminal tasks
4.0agents and tool useactive
Score in %
Definition
- Version
- 4.0
- Category
- agents and tool use
- Maintainer
- Terminal-Bench team (Stanford, Harbor and Laude Institute), USA
- Status
- active
- Home page
- https://www.tbench.ai/leaderboard
Sources
- Anthropic releases Claude Opus 5.5 with Fable 5.1-level performance at lower cost - AI News vpsranking.com https://vpsranking.com/news/ai/ai-2026-09-22-anthropic-claude-opus-5-5/
- Best AI models, October 2026: Opus 5.5, GPT-6, Gemini slash-digital.io https://slash-digital.io/en/insights/ai-models-2026/
- Claude Opus 5.5: Fable-Class Work, Cheaper to Run llm-stats.com https://llm-stats.com/blog/research/claude-opus-5-5-launch
- TERMINAL-BENCH tbench.ai https://www.tbench.ai/leaderboard
- Terminal-Bench 2.0 Leaderboard & Scores , October 2026 benchlm.ai https://www.benchlm.ai/benchmarks/terminal-bench-2
- AI Model & API Providers Analysis | Artificial Analysis artificialanalysis.ai https://artificialanalysis.ai/
- Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces arxiv.org https://arxiv.org/abs/2601.11868
Scores at a Glance
Score in %
Higher is better for this unit.
1
Claude Opus 5.5Anthropicofficial boardClaude Sonnet 5.5AnthropicClaude Fable 5.1AnthropicVals.ai, max effort
GPT-6 AstraOpenAI57.9% to 58.2% across sourcesClaude Opus 5Anthropic
Grok 4.7SpaceXAI (xAI)
DeepSeek-V4.1-FlashDeepSeek
SWE-2Cognition (owns Windsurf)
Claude Opus 5.5Anthropicofficial board64.8 %
maintainer-run2Claude Sonnet 5.5Anthropic61.8 %
vendor-reported3Claude Fable 5.1AnthropicVals.ai, max effort58.08 %
independent4GPT-6 AstraOpenAI57.9% to 58.2% across sources57.9 %
vendor-reported5Claude Opus 5Anthropic52.3 %
vendor-reported6Grok 4.7SpaceXAI (xAI)37.6 %
vendor-reported7DeepSeek-V4.1-FlashDeepSeek31.2 %
vendor-reported8SWE-2Cognition (owns Windsurf)27.3 %
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 |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 66.4 % | vendor-reported | xhigh, with safeguards, SE about 2.6 | – | Sep 2026 | vpsranking.com |
| Claude Opus 5.5 | 64.8 % | maintainer-run | official board | – | – | slash-digital.io |
| Claude Sonnet 5.5 | 61.8 % | vendor-reported | – | – | – | slash-digital.io |
| Claude Fable 5.1 | 58.08 % ±3.31 | independent | Vals.ai, max effort | – | – | vals.ai |
| GPT-6 Astra | 57.9 % | vendor-reported | 57.9% to 58.2% across sources | – | – | vpsranking.com |
| Claude Opus 5 | 52.3 % | vendor-reported | – | – | – | llm-stats.com |
| Grok 4.7 | 37.6 % | vendor-reported | – | rank 10, – | – | slash-digital.io |
| DeepSeek-V4.1-Flash | 31.2 % | vendor-reported | – | – | – | huggingface.co |
 MiMo-V2.6-Flash | 28.8 % | vendor-reported | – | MiMo-V2.6 model card comparison table | 21 Sep 2026 | huggingface.co |
| SWE-2 | 27.3 % | vendor-reported | – | – | – | orcarouter.ai |
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-4-0/, records of October 2026.