真实软件工程实现与修复任务。
Raw Evidence
编码 Agent 原始证据
完整展开当前 57 组标准化公开结果。先筛选，再检查三项原始基准、成本、时间、步骤和 Token 组成。
来源Artificial Analysis Coding Agents
指数版本v1.2
记录数57
快照2026-08-06
ALL 57 RESULTS
完整数据表
移动端保留横向滚动；首列固定，避免模型身份在检查指标时丢失。
| # | Agent / 模型 | 模型机构 | 综合指数 | DeepSWE | Terminal-Bench v2 | SWE-Atlas-QnA | 成本 / 任务 | 执行时间 | 平均步骤 | 输入 Token | 缓存 Token | 输出 Token | 总 Token | 缓存命中率 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude CodeOpus 5 (xhigh) | Anthropic | 68.1 | 60.5 | — | 54.8 | US$8.17 | 23.7 分钟 | 152 | 10,747,526 | 10,578,913 | 73,273 | 21,565,879 | 97.4% |
| 2 | Claude CodeFable 5 (max) | Anthropic | 67.2 | 66.1 | — | 48.9 | US$11.69 | 23.5 分钟 | 137 | 6,901,095 | 6,741,270 | 74,265 | 13,868,660 | 96.2% |
| 3 | Claude CodeOpus 5 (max) | Anthropic | 67.0 | 63.1 | — | 49.2 | US$8.94 | 24.2 分钟 | 166 | 11,810,978 | 11,635,411 | 81,385 | 23,699,428 | 97.5% |
| 4 | Claude CodeOpus 5 (high) | Anthropic | 65.6 | 60.8 | — | 49.2 | US$3.92 | 14.0 分钟 | 95 | 4,937,399 | 4,847,441 | 37,331 | 9,911,366 | 97.0% |
| 5 | CodexGPT-5.6 Sol (max) | OpenAI | 65.1 | 68.7 | — | 43.3 | US$6.42 | 10.2 分钟 | 112 | 6,786,107 | 6,347,957 | 35,184 | 13,169,249 | 90.3% |
| 6 | CodexGPT-5.6 Sol (high) | OpenAI | 64.1 | 64.9 | — | 45.4 | US$3.85 | 6.2 分钟 | 85 | 4,136,516 | 3,873,765 | 19,998 | 8,030,279 | 90.6% |
| 7 | Grok BuildGrok 4.5 (high) | xAI | 64.1 | 59.9 | — | 48.1 | US$2.44 | 15.5 分钟 | 61 | 1,837,535 | 1,735,633 | 25,699 | 3,598,868 | 92.4% |
| 8 | Claude CodeOpus 5 (medium) | Anthropic | 64.1 | 62.8 | — | 44.4 | US$3.17 | 12.2 分钟 | 83 | 3,984,816 | 3,911,287 | 30,206 | 7,999,358 | 97.0% |
| 9 | CodexGPT-5.6 Sol (xhigh) | OpenAI | 63.3 | 67.0 | — | 43.3 | US$4.80 | 7.3 分钟 | 94 | 5,095,035 | 4,765,876 | 25,685 | 9,886,596 | 90.8% |
| 10 | Kimi Code CLIKimi K3 | Moonshot AI | 62.6 | 63.7 | — | 36.6 | US$3.08 | 24.1 分钟 | 123 | 5,296,769 | 5,031,930 | 54,498 | 10,383,197 | 95.0% |
| 11 | Claude CodeOpus 4.8 (max) | Anthropic | 62.1 | 55.8 | — | 46.8 | US$7.72 | 23.1 分钟 | 165 | 8,919,830 | 8,733,089 | 89,375 | 17,920,579 | 96.6% |
| 12 | Muse CodeMuse Spark 1.2 (xhigh) | Meta | 61.6 | 58.1 | — | 45.2 | US$2.07 | 40.8 分钟 | 151 | 10,093,761 | 9,811,072 | 58,306 | 19,963,139 | 95.4% |
| 13 | CodexGPT-5.6 Sol (medium) | OpenAI | 61.6 | 64.0 | — | 40.3 | US$2.81 | 5.0 分钟 | 71 | 2,974,758 | 2,776,347 | 14,441 | 5,765,545 | 89.7% |
| 14 | Claude CodeQwen3.8 Max | Alibaba Cloud | 61.3 | 51.9 | — | 48.1 | US$3.23 | 29.9 分钟 | 160 | 6,469,207 | 5,809,875 | 63,118 | 12,503,742 | 90.0% |
| 15 | CodexGPT-5.5 (xhigh) | OpenAI | 61.0 | 64.3 | — | 36.0 | US$4.75 | 10.2 分钟 | 107 | 6,164,087 | 5,961,478 | 25,286 | 12,150,851 | 93.4% |
| 16 | CodexGPT-5.6 Terra (max) | OpenAI | 60.4 | 67.0 | — | 36.0 | US$2.70 | 8.2 分钟 | 96 | 4,933,734 | 4,659,332 | 37,445 | 9,630,512 | 91.8% |
| 17 | OpencodeGemini 3.7 Flash (high) | 59.6 | 57.2 | — | 30.6 | US$1.27 | 8.8 分钟 | 83 | 9,497,213 | 8,946,811 | 49,099 | 18,493,123 | 86.3% | |
| 18 | Claude CodeOpus 5 (low) | Anthropic | 59.4 | 56.9 | — | 39.2 | US$2.30 | 10.0 分钟 | 64 | 2,618,171 | 2,544,838 | 22,633 | 5,258,606 | 96.4% |
| 19 | Claude CodeOpus 4.8 (xhigh) | Anthropic | 59.1 | 51.3 | — | 42.7 | US$5.67 | 17.8 分钟 | 136 | 6,783,846 | 6,660,078 | 62,542 | 13,628,306 | 96.9% |
| 20 | OpencodeMuse Spark 1.2 (xhigh) | Meta | 58.9 | 52.8 | — | 44.1 | US$1.91 | 17.7 分钟 | 66 | 8,360,785 | 7,942,746 | 47,157 | 16,350,688 | 95.0% |
| 21 | Claude CodeOpus 4.8 (high) | Anthropic | 57.6 | 51.6 | — | 38.7 | US$3.78 | 12.7 分钟 | 105 | 4,574,602 | 4,491,206 | 40,683 | 9,188,691 | 96.8% |
| 22 | CodexGPT-5.6 Luna (max) | OpenAI | 57.2 | 63.4 | — | 32.8 | US$1.51 | 8.0 分钟 | 115 | 8,170,929 | 7,751,236 | 41,094 | 15,963,259 | 91.8% |
| 23 | Antigravity SDK v0.1.8Gemini 3.7 Flash (high) | 56.6 | 56.0 | — | 27.2 | US$1.40 | 6.4 分钟 | 111 | 7,938,344 | 7,110,911 | 64,430 | 15,113,685 | 86.8% | |
| 24 | CodexGPT-5.6 Terra (xhigh) | OpenAI | 56.0 | 58.4 | — | 32.5 | US$1.88 | 6.7 分钟 | 75 | 3,404,440 | 3,201,201 | 26,012 | 6,631,652 | 91.0% |
| 25 | Claude CodeOpus 4.8 (medium) | Anthropic | 55.8 | 49.3 | — | 36.0 | US$3.30 | 12.0 分钟 | 93 | 3,898,584 | 3,813,114 | 34,668 | 7,827,354 | 96.4% |
| 26 | CodexGPT-5.5 (medium) | OpenAI | 55.3 | 56.6 | — | 30.6 | US$2.65 | 6.4 分钟 | 79 | 3,492,715 | 3,379,606 | 13,239 | 6,885,561 | 94.1% |
| 27 | CodexGPT-5.6 Sol (low) | OpenAI | 55.2 | 53.4 | — | 34.4 | US$1.66 | 3.5 分钟 | 54 | 1,658,130 | 1,533,143 | 8,963 | 3,200,236 | 88.3% |
| 28 | OpencodeMuse Spark 1.1 (xhigh) | Meta | 54.9 | 54.3 | — | 33.3 | US$1.44 | 12.7 分钟 | 56 | 6,284,599 | 5,970,369 | 34,559 | 12,289,527 | 95.0% |
| 29 | CodexGPT-5.6 Terra (high) | OpenAI | 54.6 | 60.5 | — | 31.2 | US$1.56 | 6.0 分钟 | 67 | 2,877,096 | 2,702,277 | 20,903 | 5,600,276 | 90.5% |
| 30 | CodexGPT-5.6 Luna (xhigh) | OpenAI | 53.0 | 56.6 | — | 31.5 | US$1.21 | 6.9 分钟 | 96 | 6,575,105 | 6,200,664 | 30,840 | 12,806,609 | 91.0% |
| 31 | Devin CLISWE-1.7 Lightning Max | Cognition | 52.2 | 40.4 | — | 36.8 | US$8.52 | 10.6 分钟 | 87 | 7,834,232 | 7,700,523 | 38,917 | 15,573,673 | 97.5% |
| 32 | CodexGPT-5.6 Luna (high) | OpenAI | 51.7 | 53.4 | — | 29.0 | US$0.92 | 5.7 分钟 | 84 | 4,959,372 | 4,659,768 | 22,231 | 9,641,370 | 90.3% |
| 33 | Claude CodeOpus 4.7 (max) | Anthropic | 51.6 | 40.1 | — | 37.1 | US$5.92 | 15.8 分钟 | 108 | 8,042,458 | 7,876,338 | 47,224 | 16,129,723 | 96.7% |
| 34 | OpencodeOpus 4.7 (medium) | Anthropic | 51.3 | 39.5 | — | 36.0 | US$2.94 | 12.5 分钟 | 54 | 3,744,701 | 3,744,641 | 25,944 | 7,582,747 | 95.8% |
| 35 | CodexDeepSeek V4 Flash 0731 (max) | DeepSeek | 49.8 | 42.8 | — | 38.7 | US$0.06 | 14.5 分钟 | 106 | 10,443,572 | 10,342,245 | 58,411 | 20,844,228 | 98.1% |
| 36 | Claude CodeOpus 4.8 (low) | Anthropic | 49.0 | 41.0 | — | 28.2 | US$2.18 | 8.6 分钟 | 68 | 2,595,805 | 2,542,895 | 23,270 | 5,214,050 | 96.6% |
| 37 | CodexGPT-5.6 Terra (medium) | OpenAI | 48.0 | 45.7 | — | 28.8 | US$0.88 | 4.0 分钟 | 51 | 1,632,285 | 1,520,466 | 11,856 | 3,164,607 | 89.4% |
| 38 | OpencodeGemini 3.6 Flash (high) | 47.2 | 41.3 | — | 21.8 | US$2.08 | 10.4 分钟 | 65 | 6,765,842 | 6,211,554 | 42,763 | 13,020,160 | 83.8% | |
| 39 | Cursor CLIGPT-5.5 (medium) | OpenAI | 47.1 | 37.2 | — | 27.7 | US$2.00 | 6.4 分钟 | 78 | 2,043,913 | 1,900,873 | 11,032 | 3,955,818 | 89.2% |
| 40 | Cursor CLIOpus 4.7 (medium) | Anthropic | 46.7 | 31.6 | — | 33.9 | US$2.72 | 13.8 分钟 | 86 | 2,855,520 | 2,776,385 | 17,872 | 5,728,865 | 95.4% |
| 41 | CodexGPT-5.6 Sol (none) | OpenAI | 43.4 | 35.4 | — | 34.4 | US$1.40 | 3.3 分钟 | 55 | 1,731,981 | 1,665,241 | 7,690 | 3,404,912 | 89.2% |
| 42 | Claude CodeGLM-5.2 | Z.ai | 43.3 | 28.6 | — | 29.0 | US$6.66 | 25.1 分钟 | 127 | 5,529,736 | 1,061,226 | 28,568 | 6,619,530 | 33.0% |
| 43 | CodexDeepSeek V4 Pro 0813 (max) | DeepSeek | 42.8 | 16.2 | — | 49.2 | US$0.09 | 12.4 分钟 | 66 | 6,378,206 | 6,297,672 | 37,874 | 12,713,751 | 97.8% |
| 44 | Claude CodeOpus 4.7 (medium) | Anthropic | 42.4 | 27.4 | — | 22.6 | US$1.80 | 6.7 分钟 | 42 | 2,310,269 | 2,240,468 | 17,168 | 4,636,275 | 95.8% |
| 45 | CodexGPT-5.6 Luna (medium) | OpenAI | 42.0 | 36.6 | — | 27.2 | US$0.44 | 3.2 分钟 | 57 | 2,221,163 | 2,054,502 | 11,663 | 4,287,328 | 88.3% |
| 46 | Claude CodeSonnet 4.6 (medium) | Anthropic | 38.6 | 28.9 | — | 19.6 | US$2.04 | 13.8 分钟 | 67 | 4,205,087 | 4,083,419 | 40,070 | 8,444,053 | 94.9% |
| 47 | CodexGPT-5.6 Terra (low) | OpenAI | 38.6 | 29.8 | — | 22.8 | US$0.49 | 2.6 分钟 | 38 | 861,642 | 789,150 | 6,854 | 1,657,647 | 87.7% |
| 48 | Cursor CLIComposer 2.5 | Cursor | 38.3 | 15.9 | — | 31.2 | US$0.09 | 10.4 分钟 | 117 | 1,872,462 | 1,818,058 | 18,454 | 3,705,076 | 105.3% |
| 49 | Cursor CLIComposer 2.5 Fast | Cursor | 38.3 | 15.9 | — | 31.2 | US$0.56 | 7.9 分钟 | 117 | 2,153,034 | 2,095,608 | 20,146 | 4,259,887 | 93.6% |
| 50 | Claude CodeQwen3.7 Plus (thinking) | Alibaba Cloud | 38.1 | 19.2 | — | 23.7 | US$6.30 | 10.7 分钟 | 146 | 4,715,248 | 4,036,776 | 32,611 | 8,784,635 | 81.1% |
| 51 | Claude CodeGLM-5.1 | Z.ai | 36.9 | 18.6 | — | 24.7 | US$4.28 | 19.0 分钟 | 174 | 13,139,447 | 12,572,555 | 49,883 | 25,761,885 | 88.9% |
| 52 | Claude CodeKimi K2.6 | Moonshot AI | 33.9 | 16.5 | — | 15.9 | US$1.22 | 40.7 分钟 | 134 | 5,891,966 | 5,731,631 | 36,930 | 11,660,528 | 95.5% |
| 53 | Gemini CLIGemini 3.1 Pro (high) | 32.9 | 14.2 | — | 8.6 | US$1.04 | 11.1 分钟 | 31 | 2,465,902 | 2,248,925 | 12,978 | 4,727,804 | 86.5% | |
| 54 | Claude CodeDeepSeek V4 Pro (high) | DeepSeek | 32.8 | 8.6 | — | 19.9 | US$0.25 | 17.9 分钟 | 128 | 5,181,245 | 4,718,945 | 41,976 | 9,942,166 | 85.4% |
| 55 | CodexGPT-5.6 Luna (low) | OpenAI | 25.1 | 10.3 | — | 15.1 | US$0.19 | 1.7 分钟 | 34 | 762,149 | 667,908 | 5,609 | 1,435,666 | 84.3% |
| 56 | CodexGPT-5.6 Terra (none) | OpenAI | 23.1 | 13.3 | — | 18.5 | US$0.36 | 1.5 分钟 | 34 | 582,184 | 518,484 | 5,133 | 1,105,802 | 84.8% |
| 57 | CodexGPT-5.6 Luna (none) | OpenAI | 19.1 | 6.5 | — | 17.5 | US$0.33 | 2.2 分钟 | 54 | 1,835,573 | 1,706,450 | 7,935 | 3,549,957 | 87.5% |
“—”表示来源没有公开可核验数据。Token 为每项任务的平均用量；厂商受控实验不插入这张标准化数据表。
多步骤终端与工具调用工作流。
代码库理解与技术问答。