Agentic coding · %
Terminal-Bench 2.0
Hard, realistic tasks an agent must finish in a sandboxed terminal (Stanford × Laude Institute).
Active Contamination risk: Low 202 results · 45 models Updated Oct 22, 2026 Newest result: 2026-05-14 Official leaderboard ↗
What it measures
Task accuracy for an agent + model pair. Data compiled by Epoch AI from tbench.ai.
Limitations
Agent harness matters as much as the model; the same model appears several times with different agents.
Contamination risk: Low
Tasks carry a canary string.
Leaderboard · Subset / setting: compiled
5 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-5.3 Codex | OpenAI | 70.3% | 1 | Terminal-Bench 2.0; agent: CodeBrain-1 | Third-party compilation via Epoch AI | Feb 10, 2026 |
| 2 | Gemini 3 Pro | 62.2% | 2 | Terminal-Bench 2.0; agent: CodeBrain-1 | Third-party compilation via Epoch AI | Feb 5, 2026 | |
| 3 | Claude Sonnet 4.6 | Anthropic | 53.4% | 3 | Terminal-Bench 2.0; agent: Simplai Agent; effort: no thinking | Third-party compilation via Epoch AI | May 14, 2026 |
| 4 | Claude Sonnet 4.5 | Anthropic | 42.7% | 4 | Terminal-Bench 2.0; agent: MAYA; effort: unknown thinking | Third-party compilation via Epoch AI | Jan 4, 2026 |
| 5 | GPT-5 mini | OpenAI | 34.8% | 5 | Terminal-Bench 2.0; agent: spoox-m; effort: unknown thinking | Third-party compilation via Epoch AI | Dec 24, 2025 |
Leaderboard · Subset / setting: leaderboard
45 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-5.5 | OpenAI | 84.7% | 1 | Terminal-Bench 2.0; agent: NexAU-AHE; effort: unknown thinking | Official leaderboard via Epoch AI | Apr 23, 2026 |
| 2 | GPT-5.4 | OpenAI | 81.8% | 5 | Terminal-Bench 2.0; agent: ForgeCode; effort: unknown thinking | Official leaderboard via Epoch AI | Mar 12, 2026 |
| 3 | Gemini 3.1 Pro | 80.2% | 6 | Terminal-Bench 2.0; agent: TongAgents | Official leaderboard via Epoch AI | Feb 19, 2026 | |
| 4 | Claude Opus 4.7 | Anthropic | 80.2% | 6 | Terminal-Bench 2.0; agent: WOZCODE; effort: unknown | Official leaderboard via Epoch AI | Apr 16, 2026 |
| 5 | Claude Opus 4.6 | Anthropic | 79.8% | 8 | Terminal-Bench 2.0; agent: ForgeCode; effort: unknown thinking | Official leaderboard via Epoch AI | Mar 12, 2026 |
| 6 | GPT-5.3 Codex | OpenAI | 78.4% | 9 | Terminal-Bench 2.0; agent: SageAgent | Official leaderboard via Epoch AI | Feb 5, 2026 |
| 7 | Gemini 3 Pro | 69.4% | 26 | Terminal-Bench 2.0; agent: Ante | Official leaderboard via Epoch AI | Nov 18, 2025 | |
| 8 | GPT-5.2 Codex | OpenAI | 66.5% | 29 | Terminal-Bench 2.0; agent: Deep Agents | Official leaderboard via Epoch AI | Jan 14, 2026 |
| 9 | GPT-5.2 | OpenAI | 64.9% | 33 | Terminal-Bench 2.0; agent: Droid; effort: unknown thinking | Official leaderboard via Epoch AI | Dec 11, 2025 |
| 10 | Gemini 3 Flash | 64.3% | 36 | Terminal-Bench 2.0; agent: Junie CLI | Official leaderboard via Epoch AI | Dec 17, 2025 | |
| 11 | Claude Opus 4.5 | Anthropic | 63.1% | 37 | Terminal-Bench 2.0; agent: Droid; effort: unknown thinking | Official leaderboard via Epoch AI | Nov 24, 2025 |
| 12 | GPT-5.1-Codex-mini | OpenAI | 61.6% | 44 | Terminal-Bench 2.0; agent: hookele | Official leaderboard via Epoch AI | Nov 12, 2025 |
| 13 | GPT-5.1-Codex-Max | OpenAI | 60.5% | 48 | Terminal-Bench 2.0; agent: Codex CLI | Official leaderboard via Epoch AI | Nov 19, 2025 |
| 14 | GPT-5.1-Codex | OpenAI | 57.8% | 53 | Terminal-Bench 2.0; agent: Codex CLI | Official leaderboard via Epoch AI | Nov 16, 2025 |
| 15 | Grok 4.20 | xAI | 57.3% | 56 | Terminal-Bench 2.0; agent: Grok CLI | Official leaderboard via Epoch AI | Mar 10, 2026 |
| 16 | Claude Sonnet 4.6 | Anthropic | 53.4% | 63 | Terminal-Bench 2.0; agent: Simplai Agent; effort: unknown thinking | Official leaderboard via Epoch AI | Feb 17, 2026 |
| 17 | GLM-5OW | Zhipu AI (Z.ai) | 52.4% | 64 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Feb 11, 2026 |
| 18 | GPT-5 | OpenAI | 49.6% | 70 | Terminal-Bench 2.0; agent: Codex CLI; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| 19 | GPT-5.1 | OpenAI | 47.6% | 72 | Terminal-Bench 2.0; agent: Terminus 2; effort: medium | Official leaderboard via Epoch AI | Nov 16, 2025 |
| 20 | Claude Sonnet 4.5 | Anthropic | 46.5% | 76 | Terminal-Bench 2.0; agent: CAMEL-AI; effort: unknown thinking | Official leaderboard via Epoch AI | Sep 29, 2025 |
| 21 | MiniMax-M2.7OW | MiniMax | 45.1% | 77 | Terminal-Bench 2.0; agent: IndusAGI Coding Agent | Official leaderboard via Epoch AI | Mar 18, 2026 |
| 22 | GPT-5-Codex | OpenAI | 44.3% | 78 | Terminal-Bench 2.0; agent: Codex CLI | Official leaderboard via Epoch AI | Sep 15, 2025 |
| 23 | Kimi K2.5OW | Moonshot AI | 43.2% | 84 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Jan 27, 2026 |
| 24 | MiniMax-M2.5OW | MiniMax | 42.7% | 91 | Terminal-Bench 2.0; agent: cchuter | Official leaderboard via Epoch AI | Feb 12, 2026 |
| 25 | DeepSeek-V3.2OW | DeepSeek | 39.5% | 102 | Terminal-Bench 2.0; agent: Terminus 2; effort: Thinking; Novita | Official leaderboard via Epoch AI | Dec 1, 2025 |
| 26 | Claude Opus 4.1 | Anthropic | 38.0% | 103 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Oct 31, 2025 |
| 27 | MiniMax-M2.1OW | MiniMax | 36.6% | 109 | Terminal-Bench 2.0; agent: Crux | Official leaderboard via Epoch AI | Dec 23, 2025 |
| 28 | Kimi K2 ThinkingOW | Moonshot AI | 35.7% | 110 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Nov 6, 2025 |
| 29 | Claude Haiku 4.5 | Anthropic | 35.5% | 112 | Terminal-Bench 2.0; agent: Goose; effort: unknown thinking | Official leaderboard via Epoch AI | Oct 15, 2025 |
| 30 | GPT-5 mini | OpenAI | 34.8% | 117 | Terminal-Bench 2.0; agent: spoox-o-m; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| 31 | GLM-4.7OW | Zhipu AI (Z.ai) | 33.4% | 122 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Dec 22, 2025 |
| 32 | Gemini 2.5 Pro (Jun 2025) | 32.6% | 124 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Jun 17, 2025 | |
| 33 | MiniMax-M2OW | MiniMax | 30.0% | 128 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Oct 27, 2025 |
| 34 | Kimi K2 (Jul 2025)OW | Moonshot AI | 27.8% | 136 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Nov 1, 2025 |
| 35 | Grok 4 | xAI | 27.2% | 140 | Terminal-Bench 2.0; agent: OpenHands | Official leaderboard via Epoch AI | Nov 2, 2025 |
| 36 | Qwen3-Coder-480B-A35BOW | Alibaba (Qwen) | 27.2% | 141 | Terminal-Bench 2.0; agent: Dakou Agent | Official leaderboard via Epoch AI | Jul 22, 2025 |
| 37 | grok-code-fast-1 | xAI | 25.8% | 147 | Terminal-Bench 2.0; agent: Mini-SWE-Agent | Official leaderboard via Epoch AI | Aug 28, 2025 |
| 38 | Qwen 3.6 35B-A3BOW | Alibaba (Qwen) | 24.6% | 153 | Terminal-Bench 2.0; agent: little-coder | Official leaderboard via Epoch AI | Apr 16, 2026 |
| 39 | GLM-4.6OW | Zhipu AI (Z.ai) | 24.5% | 154 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Sep 30, 2025 |
| 40 | GPT-5 nano | OpenAI | 21.8% | 164 | Terminal-Bench 2.0; agent: spoox-o-m; effort: unknown thinking | Official leaderboard via Epoch AI | Aug 7, 2025 |
| 41 | gpt-oss-120bOW | OpenAI | 18.7% | 167 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Nov 1, 2025 |
| 42 | Gemini 2.5 Flash (Sep 2025) | 17.1% | 169 | Terminal-Bench 2.0; agent: Mini-SWE-Agent | Official leaderboard via Epoch AI | Nov 3, 2025 | |
| 43 | Gemini 2.5 Flash (Jun 2025) | 17.1% | 170 | Terminal-Bench 2.0; agent: Mini-SWE-Agent | Official leaderboard via Epoch AI | Jun 17, 2025 | |
| 44 | Qwen3.5-9BOW | Alibaba (Qwen) | 9.2% | 189 | Terminal-Bench 2.0; agent: little-coder | Official leaderboard via Epoch AI | Mar 2, 2026 |
| 45 | gpt-oss-20bOW | OpenAI | 3.4% | 194 | Terminal-Bench 2.0; agent: Mini-SWE-Agent | Official leaderboard via Epoch AI | Nov 3, 2025 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2025-06-1732.6%Gemini 2.5 Pro (Jun 2025)
- 2025-08-0538.0%Claude Opus 4.1
- 2025-08-0749.6%GPT-5
- 2025-11-1261.6%GPT-5.1-Codex-mini
- 2025-11-1869.4%Gemini 3 Pro
- 2026-02-0578.4%GPT-5.3 Codex
- 2026-02-1980.2%Gemini 3.1 Pro
- 2026-03-1281.8%GPT-5.4
- 2026-04-2384.7%GPT-5.5
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| Epoch AI Benchmarking Hub Epoch AI, 'Capabilities & benchmarking', https://epoch.ai/benchmarks (CC BY 4.0) | CC BY 4.0 | Open licence – full data with attribution | Oct 22, 2026 |
FAQ
What does Terminal-Bench 2.0 measure?
Hard, realistic tasks an agent must finish in a sandboxed terminal (Stanford × Laude Institute). Task accuracy for an agent + model pair. Data compiled by Epoch AI from tbench.ai.
How reliable is Terminal-Bench 2.0?
Agent harness matters as much as the model; the same model appears several times with different agents. Contamination risk: Low – Tasks carry a canary string.