代理式程式設計
藉助工具迴圈解決真實程式碼儲存庫和終端機任務。這裡每個分數衡量的都是「模型 + 框架」的整體，所以同一個模型可能佔兩行，分別對應不同框架，這不是重複。 按它在多少項基準裡進入前 10 排序，一樣多時看最好名次。每列保持各自的排名，不做平均。 「—」表示該榜沒有收錄這個模型，不是零分。尚未識別出具體模型的條目不在此處出現，但仍顯示在各自的原始榜單上。
| 模型 | 進入前 10 | Artificial Analysis LLM Leaderboard Terminal-Bench 4.0 | Artificial Analysis Coding Agents Artificial Analysis Coding Agent Index | Artificial Analysis Coding Agents Terminal-Bench v4 | HarnessTax — SWE-bench Lite SWE-bench Lite | HarnessTax — Terminal-Bench 2.0 Terminal-Bench 2.0 | Terminal-Bench 4.0 Terminal-Bench 4.0 |
|---|---|---|---|---|---|---|---|
 Claude Sonnet 5.5 Anthropic | 4/6 | #1 63.6% | #1 0.684 | #1 66.2% | — | — | #2 61.8% |
| Claude Opus 5.5 Anthropic | 4/6 | #2 59.6% | #2 0.660 | #2 63.1% | — | — | #1 64.8% |
| Claude Fable 5 Anthropic | 4/6 | #9 42.4% | — | — | #1 97.8% | #3 75.6% | #8 44.5% |
 GPT 5.6 Sol OpenAI | 4/6 | #11 39.9% | #10 0.546 | #10 37.4% | #3 77.8% | #1 83.3% | #11 37.3% |
| GPT 6 Astra OpenAI | 4/6 | #3 59.6% | #6 0.616 | #5 55.6% | — | — | #3 58.2% |
| Claude Fable 5.1 Anthropic | 4/6 | #6 55.1% | #5 0.622 | #3 57.6% | — | — | #5 57.9% |
| GPT 6.1 Sol OpenAI | 4/6 | #5 56.1% | #4 0.629 | #6 54.5% | — | — | #4 58.2% |
| Claude Opus 5 Anthropic | 4/6 | #7 49.0% | #7 0.597 | #7 54.5% | — | — | #6 53.9% |
| GPT 6 Sol OpenAI | 4/6 | #8 43.9% | #8 0.567 | #8 43.4% | — | — | #7 49.4% |
 Gemini 4 Argon Google | 3/6 | #4 57.1% | #3 0.638 | #4 56.1% | — | — | — |
 GLM 5.3 Z AI | 3/6 | #10 41.9% | #12 0.536 | #9 39.9% | — | — | #9 41.8% |
| Claude Opus 4.8 Anthropic | 2/6 | #25 21.7% | — | — | #2 88.9% | #5 72.2% | #14 23.6% |
| GPT 5.6 Luna OpenAI | 2/6 | #37 11.6% | #17 0.432 | #19 14.6% | #7 55.6% | #2 76.7% | #18 17.3% |
 Kimi K3 Kimi | 2/6 | #34 12.6% | #13 0.519 | #15 21.2% | #4 76.7% | #4 73.3% | — |
| Claude Sonnet 4.6 Anthropic | 2/6 | #45 3.0% | — | — | #5 68.9% | #6 65.6% | — |
| Claude Haiku 4.5 Anthropic | 2/6 | #75 0.0% | — | — | #6 60.0% | #7 47.8% | — |
 Grok 4.7 SpaceXAI | 2/6 | #22 25.8% | #9 0.563 | #11 33.3% | — | — | #10 37.6% |