Terminal-Bench 2.0
Hard tasks done entirely in a terminal: compiling, debugging, data wrangling, sysadmin. Best reported agent per model.
Agent Index weight
8%
in the Agentic coding pillar
Models tested
44
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | GPT-5.5 Supported OpenAI | 84.7% | NexAU-AHE | — | Apr 23, 2026 | |
| 2 | GPT-5.4 Supported OpenAI | 81.8% | ForgeCode | — | Mar 12, 2026 | |
| 3 | Claude Opus 4.7 Supported Anthropic | 80.2% | WOZCODE | — | Apr 16, 2026 | |
| 4 | Gemini 3.1 Pro Preview Supported Google | 80.2% | TongAgents | — | Feb 19, 2026 | |
| 5 | Claude Opus 4.6 Supported Anthropic | 79.8% | ForgeCode | — | Mar 12, 2026 | |
| 6 | GPT-5.3 Codex Supported OpenAI | 78.4% | SageAgent | — | Feb 5, 2026 | |
| 7 | Gemini 3 Pro Preview Supported Google | 69.4% | Ante | — | Nov 18, 2025 | |
| 8 | GPT-5.2 Codex OpenAI | 66.5% | Deep Agents | — | Jan 14, 2026 | |
| 9 | GPT-5.2 Supported OpenAI | 64.9% | Droid | — | Dec 11, 2025 | |
| 10 | Gemini 3 Flash Preview Supported Google | 64.3% | Junie CLI | — | Dec 17, 2025 | |
| 11 | Claude Opus 4.5 (latest) Supported Anthropic | 63.1% | Droid | — | Nov 24, 2025 | |
| 12 | GPT-5.1 Codex mini OpenAI | 61.6% | hookele | — | Nov 12, 2025 | |
| 13 | GPT-5.1 Codex Max Estimated OpenAI | 60.4% | Codex CLI | — | Nov 19, 2025 | |
| 14 | GPT-5.1 Codex OpenAI | 57.8% | Codex CLI | — | Nov 16, 2025 | |
| 15 | Grok 4.20 (Non-Reasoning) Estimated xAI | 57.3% | Grok CLI | — | Mar 10, 2026 | |
| 16 | Claude Sonnet 4.6 Supported Anthropic | 53.4% | Simplai Agent | — | Feb 17, 2026 | |
| 17 | GLM-5 Supported Z.ai (Zhipu) | 52.4% | Terminus 2 | — | Feb 11, 2026 | |
| 18 | GPT-5 Supported OpenAI | 49.6% | Codex CLI | — | Aug 7, 2025 | |
| 19 | GPT-5.1 Supported OpenAI | 47.6% | Terminus 2 | — | Nov 16, 2025 | |
| 20 | Claude Sonnet 4.5 (latest) Supported Anthropic | 46.5% | CAMEL-AI | — | Sep 29, 2025 | |
| 21 | MiniMax-M2.7 MiniMax | 45.1% | IndusAGI Coding Agent | — | Mar 18, 2026 | |
| 22 | GPT-5-Codex OpenAI | 44.3% | Codex CLI | — | Sep 15, 2025 | |
| 23 | Kimi K2.5 Supported Moonshot AI | 43.2% | Terminus 2 | — | Jan 27, 2026 | |
| 24 | MiniMax-M2.5 Estimated MiniMax | 42.7% | cchuter | — | Feb 12, 2026 | |
| 25 | DeepSeek V3.2 Supported DeepSeek | 39.6% | Terminus 2 | — | Dec 1, 2025 | |
| 26 | Claude Opus 4.1 (latest) Supported Anthropic | 38.0% | Terminus 2 | — | Oct 31, 2025 | |
| 27 | MiniMax-M2.1 MiniMax | 36.6% | Crux | — | Dec 23, 2025 | |
| 28 | Kimi K2 Thinking Estimated Moonshot AI | 35.7% | Terminus 2 | — | Nov 6, 2025 | |
| 29 | Claude Haiku 4.5 Estimated Anthropic | 35.5% | Goose | — | Oct 15, 2025 | |
| 30 | GPT-5 Mini Estimated OpenAI | 34.8% | spoox-o-m | — | Aug 7, 2025 | |
| 31 | GLM-4.7 Estimated Z.ai (Zhipu) | 33.4% | Terminus 2 | — | Dec 22, 2025 | |
| 32 | Gemini 2.5 Pro Supported Google | 32.6% | Terminus 2 | — | Jun 17, 2025 | |
| 33 | MiniMax-M2 Estimated MiniMax | 30.0% | Terminus 2 | — | Oct 27, 2025 | |
| 34 | Kimi K2 0905 Estimated Moonshot AI | 27.8% | Terminus 2 | — | Nov 1, 2025 | |
| 35 | Grok 4 Supported xAI | 27.2% | OpenHands | — | Nov 2, 2025 | |
| 36 | Qwen3 Coder 480B A35B Estimated Alibaba (Qwen) | 27.2% | Dakou Agent | — | Jul 22, 2025 | |
| 37 | Grok Code Fast 1 xAI | 25.8% | Mini-SWE-Agent | — | Aug 28, 2025 | |
| 38 | Qwen3.6 35B-A3B Alibaba (Qwen) | 24.6% | little-coder | — | Apr 16, 2026 | |
| 39 | GLM-4.6 Z.ai (Zhipu) | 24.5% | Terminus 2 | — | Sep 30, 2025 | |
| 40 | GPT-5 Nano OpenAI | 21.8% | spoox-o-m | — | Aug 7, 2025 | |
| 41 | GPT OSS 120B Supported OpenAI | 18.7% | Terminus 2 | — | Nov 1, 2025 | |
| 42 | Gemini 2.5 Flash Estimated Google | 17.1% | Mini-SWE-Agent | — | Nov 3, 2025 | |
| 43 | Qwen3.5 9B Alibaba (Qwen) | 9.2% | little-coder | — | Mar 2, 2026 | |
| 44 | GPT OSS 20B OpenAI | 3.4% | Mini-SWE-Agent | — | Nov 3, 2025 |
For models without a Terminal-Bench 2.0 result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 31 models (residual SD 0.38 on the logit scale). Estimates are labelled and widen the model’s range.