# Terminal-Bench 2.0 > A benchmark for agentic software engineering tasks executed in real terminal environments. Models must inspect files, run commands, edit code, and recover from errors over multi-step workflows. Canonical page: https://benchlm.ai/benchmarks/terminal-bench-2 - Category: [Agentic](/agentic) - Last updated: September 10, 2026 Latest rankings: [the latest benchmark rankings](/benchmarks/terminal-bench-4) ## About Terminal-Bench 2.0 - Year: 2026 - Tasks: Terminal-based software tasks - Format: Interactive CLI agent evaluation - Difficulty: Professional software engineering - Paper: [Terminal-Bench 2.0](https://www.tbench.ai/) Terminal-Bench 2.0 focuses on realistic CLI and repository workflows rather than toy code generation. It is a strong proxy for how useful a model is inside coding agents and autonomous developer tools. Terminal-Bench 2.0 is currently weighted in BenchLM's scoring formula. The Agentic category carries 22% of the overall score, and Terminal-Bench 2.0 contributes 30% of that category score. ## Leaderboard (67 models) | Rank | Model | Creator | Score | |------|-------|---------|-------| | 1 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 91.9% | | 2 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 88.3% | | 3 | [Claude Mythos 5](/models/claude-mythos-5) | Anthropic | 88% | | 4 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 87.4% | | 5 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 84.7% | | 6 | [Claude Fable 5](/models/claude-fable) | Anthropic | 84.3% | | 7 | [Grok 4.5](/models/grok-4-5) | xAI | 83.3% | | 8 | [Sakana Fugu-Ultra](/models/sakana-fugu-ultra) | Sakana AI | 82.1% | | 9 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 82% | | 10 | [SWE-1.7](/models/swe-1-7) | Cognition | 81.5% | | 11 | [GLM-5.2](/models/glm-5-2) | Z.AI | 81% | | 12 | [Claude Sonnet 5](/models/claude-sonnet-5) | Anthropic | 80.4% | | 13 | [Sakana Fugu](/models/sakana-fugu) | Sakana AI | 80.2% | | 14 | [Muse Spark 1.1](/models/muse-spark-1-1) | Meta | 80% | | 15 | [Ornith-1.0-397B](/models/ornith-1-0-397b) | DeepReinforce AI | 77.5% | | 16 | [GPT-5.3 Codex](/models/gpt-5-3-codex) | OpenAI | 77.3% | | 17 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 76.2% | | 18 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 75.1% | | 19 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 74.6% | | 20 | [Qwen3.7 Plus](/models/qwen3-7-plus) | Alibaba | 70.3% | | 21 | [Laguna S 2.1](/models/laguna-s-2-1) | Poolside | 70.2% | | 22 | [Qwen3.7 Max](/models/qwen3-7-max) | Alibaba | 69.7% | | 23 | [Claude Opus 4.7 (Adaptive)](/models/claude-opus-4-7-adaptive) | Anthropic | 69.4% | | 24 | [Composer 2.5](/models/composer-2-5) | Cursor | 69.3% | | 25 | [MiMo-V2.5-Pro](/models/mimo-v2-5-pro) | Xiaomi | 68.4% | | 26 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | DeepSeek | 67.9% | | 27 | [Kimi K2.6](/models/kimi-2-6) | Moonshot AI | 66.7% | | 28 | [MiniMax M3](/models/minimax-m3) | MiniMax | 66% | | 29 | [MiMo-V2.5](/models/mimo-v2-5) | Xiaomi | 65.8% | | 30 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 65.4% | | 31 | [Qwen 3.6 Max (preview)](/models/qwen3-6-max-preview) | Alibaba | 65.4% | | 32 | [Inkling-Small](/models/inkling-small) | Thinking Machines Lab | 64.7% | | 33 | [Ornith-1.0-35B](/models/ornith-1-0-35b) | DeepReinforce AI | 64.2% | | 34 | [Inkling](/models/inkling) | Thinking Machines Lab | 63.8% | | 35 | [GLM-5.1](/models/glm-5-1) | Z.AI | 63.5% | | 36 | [Composer 2](/models/composer-2) | Cursor | 61.7% | | 37 | [Qwen3.6 Plus](/models/qwen3-6-plus) | Alibaba | 61.6% | | 38 | [GPT-5.4 mini](/models/gpt-5-4-mini) | OpenAI | 60% | | 39 | [Step 3.7 Flash](/models/step-3-7-flash) | StepFun | 59.5% | | 40 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 59.3% | | 41 | [Qwen3.6-27B](/models/qwen3-6-27b) | Alibaba | 59.3% | | 42 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | Anthropic | 59.1% | | 43 | [Muse Spark](/models/muse-spark) | Meta | 59% | | 44 | [MiniMax M2.7](/models/minimax-m2-7) | MiniMax | 57% | | 45 | [DeepSeek V4 Flash 0731](/models/deepseek-v4-flash-0731) | DeepSeek | 56.9% | | 46 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | NVIDIA | 56.4% | | 47 | [GLM-5](/models/glm-5) | Z.AI | 56.2% | | 48 | [Hy3 Preview](/models/hy3-preview) | Tencent | 54.4% | | 49 | [Gemini 3.5 Flash-Lite](/models/gemini-3-5-flash-lite) | Google | 54% | | 50 | [Qwen3.5 397B](/models/qwen3-5-397b) | Alibaba | 52.5% | | 51 | [Qwen3.6-35B-A3B](/models/qwen3-6-35b-a3b) | Alibaba | 51.5% | | 52 | [Kimi K2.5 (Reasoning)](/models/kimi-k2-5-reasoning) | Moonshot AI | 50.8% | | 53 | [Kimi K2.5](/models/kimi-k2-5) | Moonshot AI | 50.8% | | 54 | [Claude Sonnet 4.5](/models/claude-sonnet-4-5) | Anthropic | 50% | | 55 | [Qwen3.5-122B-A10B](/models/qwen3-5-122b-a10b) | Alibaba | 49.4% | | 56 | [Grok 4.20](/models/grok-4-20-beta) | xAI | 47.1% | | 57 | [GPT-5.4 nano](/models/gpt-5-4-nano) | OpenAI | 46.3% | | 58 | [MAI-Thinking-1](/models/mai-thinking-1) | Microsoft | 46% | | 59 | [Laguna M.1](/models/laguna-m-1) | Poolside | 45.8% | | 60 | [Ornith-1.0-9B](/models/ornith-1-0-9b) | DeepReinforce AI | 43.1% | | 61 | [Qwen3.5-27B](/models/qwen3-5-27b) | Alibaba | 41.6% | | 62 | [GLM-4.7](/models/glm-4-7) | Z.AI | 41% | | 63 | [Qwen3.5-35B-A3B](/models/qwen3-5-35b-a3b) | Alibaba | 40.5% | | 64 | [Laguna XS 2.1](/models/laguna-xs-2-1) | Poolside | 37.5% | | 65 | [Laguna XS.2](/models/laguna-xs-2) | Poolside | 35.7% | | 66 | [LongCat-Flash-Lite-Sparse](/models/longcat-flash-lite-sparse) | Meituan | 33.7% | | 67 | [Nemotron 3.5 Lightning 30B A3B NVFP4](/models/nemotron-3-5-lightning-30b-a3b-nvfp4) | NVIDIA | 23.5% | ## FAQ ### What does Terminal-Bench 2.0 measure? A benchmark for agentic software engineering tasks executed in real terminal environments. Models must inspect files, run commands, edit code, and recover from errors over multi-step workflows. ### Which model scores highest on Terminal-Bench 2.0? GPT-5.6 Sol by OpenAI currently leads with a score of 91.9% on Terminal-Bench 2.0. ### How many models are evaluated on Terminal-Bench 2.0? 67 AI models have been evaluated on Terminal-Bench 2.0 on BenchLM. ### Does Terminal-Bench 2.0 affect BenchLM's overall score? Yes. Terminal-Bench 2.0 is a weighted benchmark inside the Agentic category, which carries 22% of BenchLM's overall score. Terminal-Bench 2.0 itself contributes 30% of that category score. ## Compare Top Models on Terminal-Bench 2.0 - [GPT-5.6 Sol vs Kimi K3](/compare/gpt-5-6-sol-vs-kimi-k3) - [Kimi K3 vs Claude Mythos 5](/compare/claude-mythos-5-vs-kimi-k3) - [Claude Mythos 5 vs GPT-5.6 Terra](/compare/claude-mythos-5-vs-gpt-5-6-terra) - [GPT-5.6 Terra vs GPT-5.6 Luna](/compare/gpt-5-6-luna-vs-gpt-5-6-terra) ## Related Reading - [Terminal-Bench 2.0 benchmark explainer](/blog/posts/terminal-bench-2-agentic-benchmark)