Claude Opus 5vsKimi K3
在 50 个同模式 benchmark 中，Claude Opus 5 整体领先：Claude Opus 5 领先 44 项，Kimi K3 领先 6 项，持平 0 项，平均分差 +13.23。另有 8 项测试模式不同，仅供参考。
Claude Opus 5
Anthropic · 2026-07-24 · 推理大模型
Kimi K3
Moonshot AI · 2026-07-16 · 推理大模型
Claude Opus 544 项(88%)(12%)6 项Kimi K3
评测分数
按能力类目分组，每组内按分差大小排列；共 50 项。
仓库修复与多文件工程
Claude Opus 5 领先 4/4| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| FrontierSWE v2 | 52.016 / 22Max (With Tools) | 25.8612 / 22Max (With Tools) | +26.15 |
| NL2Repo-Bench | 75.301 / 20Max (With Tools) | 588 / 20Max (With Tools) | +17.30 |
| APEX-SWE | 63.703 / 72Max (With Tools) | 4818 / 72Max (With Tools) | +15.70 |
| DeepSWE | 73.6512 / 106Max (With Tools) | 68.5134 / 106Max (With Tools) | +5.14 |
数学
Claude Opus 5 领先 4/4| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 73.1719 / 77Max (With Tools) | 39.0233 / 77Max (With Tools) | +34.15 |
| Riemann-bench | 684 / 42Max (With Tools + Internet) | 37.6017 / 42Max (With Tools + Internet) | +30.40 |
| FrontierMath v2 | 85.6112 / 120Max (With Tools) | 72.1825 / 120Max (With Tools) | +13.43 |
| ProofBench v1.1 (Lean 4) | 995 / 53Max (With Tools) | 8712 / 53Max (With Tools) | +12 |
图像感知与视觉推理
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| ZeroBench Main | 522 / 8Max (With Tools) | 416 / 8Max (With Tools) | +11 |
| BabyVision | 94.101 / 13Max (With Tools) | 85.707 / 13Max (With Tools) | +8.40 |
| MMMU-Pro | 84.708 / 233Max (No Tools) | 80.5038 / 233Max (No Tools) | +4.20 |
客服与业务流程
Claude Opus 5 领先 2/3| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| τ³-Banking | 48.716 / 170Max (With Tools) | 37.1142 / 170Max (With Tools) | +11.60 |
| Public Benefits Bench v1.1 | 76.931 / 50Max (With Tools) | 68.2712 / 50Max (With Tools) | +8.66 |
| SAGE | 49.4321 / 91Max (With Tools) | 54.264 / 91Max (With Tools) | -4.83 |
法律
Kimi K3 领先 2/3| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Legal Research Bench | 55.291 / 84Max (With Tools) | 44.2021 / 84Max (With Tools) | +11.09 |
| Harvey's Legal Agent Benchmark | 6.6721 / 51Max (With Tools) | 12.928 / 51Max (With Tools) | -6.25 |
| Harvey Lab-AA | 93.464 / 44Max (With Tools) | 94.642 / 44Max (With Tools) | -1.18 |
自主开发与终端任务
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 51.8217 / 108Max (With Tools) | 12.6062 / 108Max (With Tools) | +39.22 |
| Terminal-Bench 3.0 | 43.301 / 11Max (With Tools) | 17.708 / 11Max (With Tools) | +25.60 |
| Vibe Code Bench 1-100 | 28.533 / 25Max (With Tools) | 18.2414 / 25Max (With Tools) | +10.29 |
金融
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| EMB (Excel Modeling) | 73.566 / 77Max (With Tools) | 66.4021 / 77Max (With Tools) | +7.16 |
| Tax Agent Bench | 75.062 / 71Max (With Tools) | 68.677 / 71Max (With Tools) | +6.39 |
| Finance Agent v2 | 58.6310 / 76Max (With Tools) | 54.4027 / 76Max (With Tools) | +4.23 |
临床工作与医疗决策
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| MedCode | 63.571 / 107Max (With Tools) | 48.8827 / 107Max (With Tools) | +14.69 |
| MedScribe | 90.984 / 109Max (With Tools) | 87.9615 / 109Max (With Tools) | +3.02 |
文档与图表理解
胶着 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Chartography | 845 / 17Max (With Tools) | 68.1011 / 17Max (With Tools) | +15.90 |
| GDP.pdf | 21.6034 / 126Max (No Tools) | 2233 / 126Max (No Tools) | -0.40 |
科学知识与推理
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| CritPt | 29.1012 / 208Max (No Tools) | 23.4031 / 208Max (No Tools) | +5.70 |
| GPQA Diamond | 87.8865 / 250Thinking Low (No Tools) | 84.8591 / 250Thinking Low (No Tools) | +3.03 |
科学计算与科研编程
胶着 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 305 / 18Max (With Tools) | 7.1014 / 18Max (With Tools) | +22.90 |
| SciCode | 56.4028 / 137Max (No Tools) | 59.508 / 137Max (No Tools) | -3.10 |
程序生成与编辑
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Program Bench | 376 / 18Max (With Tools) | 17.5015 / 18Max (With Tools) | +19.50 |
| Vibe Code Bench v1.1 | 88.4010 / 114Max (With Tools) | 84.9617 / 114Max (With Tools) | +3.44 |
跨应用与工具编排
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| APEX-Agents | 65.806 / 61Max (With Tools) | 4145 / 61Max (With Tools) | +24.80 |
| Agents' Last Exam | 32.205 / 21Max (With Tools) | 28.3011 / 21Max (With Tools) | +3.90 |
专业数据分析
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 53.753 / 29Max (With Tools + Internet) | 38.7514 / 29Max (With Tools + Internet) | +15 |
事实查找与深度检索
Kimi K3 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| BrowseComp | 90.803 / 61Max (With Tools + Internet) | 91.202 / 61Max (With Tools + Internet) | -0.40 |
写作与创意表达
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Creative Writing | 2,1333 / 114Normal (No Tools) | 2,0825 / 114Normal (No Tools) | +50.30 |
办公与文档流程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| AutomationBench | 50.3013 / 148Max (With Tools) | 46.7018 / 148Max (With Tools) | +3.60 |
多轮记忆与持续上下文
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Context Arena | 97.721 / 154Max (No Tools) | 71.7567 / 154Max (No Tools) | +25.97 |
抽象归纳与泛化
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 97.5013 / 249Max (With Tools) | 94.5042 / 249Max (With Tools) | +3 |
机器学习工程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| PostTrain Bench | 358 / 17Reported best (effort unspecified) | 3212 / 17Reported best (effort unspecified) | +3 |
桌面工作流
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| OSWorld 2.0 | 70.574 / 20Max (With Tools) | 58.3016 / 20Max (With Tools) | +12.27 |
生物与基因
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| BioMysteryBench | 79.263 / 21Max (With Tools) | 72.598 / 21Max (With Tools) | +6.67 |
算法与竞赛编程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| IOI (Vals v2) | 84.338 / 45Max (With Tools) | 48.9426 / 45Max (With Tools) | +35.39 |
维护、优化与工程管理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Code Migration | 57.478 / 79Max (With Tools) | 16.1056 / 79Max (With Tools) | +41.37 |
综合偏好评测
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Text Arena (Coding) | 1,7121 / 35Max (No Tools) | 1,6822 / 35Max (No Tools) | +30.13 |
综合知识考试
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| HLE | 63.606 / 244Max (With Tools) | 59.8012 / 244Max (With Tools) | +3.80 |
跨能力综合测试
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| LiveBench 2.0 | 82.098 / 49Max (task-specific tools) | 63.3934 / 49Max (task-specific tools) | +18.70 |
跨能力聚合指数
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Kimi K3 | 分差 |
|---|---|---|---|
| Vals Index | 67.214 / 42Max (With Tools) | 57.8120 / 42Max (With Tools) | +9.40 |
测试模式不同的成绩
共 8 项。两款模型的公开成绩来自不同测试模式，分数不直接可比，不计入胜负和平均分差。
| 评测项 | Claude Opus 5 | Kimi K3 |
|---|---|---|
| Vals Index v2.1 | 63.67Max (task-specific tools) | 50.70Reported best (effort unspecified) |
| FrontierCode 1.1 Main | 53.40Reported best (effort unspecified) | 44.20Normal (With Tools) |
| SimpleBench | 80.60Thinking High (No Tools) | 60.70Max (No Tools) |
| LiveBench | 80.08Max (No Tools) | 79.19Reported best (effort unspecified) |
| WeirdML v3 | 0.19Extra-High (With Tools) | 0.08Thinking High (With Tools) |
| SWE-Bench Pro V2 | 99.40Extra-High (With Tools) | 97.70Max (With Tools) |
| SWE-Bench Pro V2 Hard | 98Extra-High (With Tools) | 88.20Max (With Tools) |
| MCP-Atlas | 85.80Extra-High (With Tools) | 82.30Max (With Tools) |
规格对比
| 字段 | Claude Opus 5 | Kimi K3 |
|---|---|---|
| 发布机构 | Anthropic | Moonshot AI |
| 发布时间 | 2026-07-24 | 2026-07-16 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | 稠密模型 | MoE 架构 |
| 参数规模 | 暂无数据 | 2.8万亿 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 128K | 1M |
API 调用价格
价格优先使用 DataLearner 配置的 API 记录；缺失项不做推测。
| 价格项 | Claude Opus 5 | Kimi K3 |
|---|---|---|
| 文本输入 | $5 / 1M tokens | ¥20 / 1M tokens |
| 文本输出 | $25 / 1M tokens | ¥100 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | ¥2 / 1M tokens |
| 缓存写入 | $6.25 / 1M tokens | ¥20 / 1M tokens |
小结
- Claude Opus 5在以下类目领先:仓库修复与多文件工程 (4/4)、数学 (4/4)、图像感知与视觉推理 (3/3)、客服与业务流程 (2/3)、自主开发与终端任务 (3/3)、金融 (3/3)、临床工作与医疗决策 (2/2)、科学知识与推理 (2/2)、程序生成与编辑 (2/2)、跨应用与工具编排 (2/2)、专业数据分析 (1/1)、写作与创意表达 (1/1)、办公与文档流程 (1/1)、多轮记忆与持续上下文 (1/1)、抽象归纳与泛化 (1/1)、机器学习工程 (1/1)、桌面工作流 (1/1)、生物与基因 (1/1)、算法与竞赛编程 (1/1)、维护、优化与工程管理 (1/1)、综合偏好评测 (1/1)、综合知识考试 (1/1)、跨能力综合测试 (1/1)、跨能力聚合指数 (1/1)
- Kimi K3在以下类目领先:法律 (2/3)、事实查找与深度检索 (1/1)
- 胶着类目:文档与图表理解、科学计算与科研编程
50 个同模式 benchmark 上，Claude Opus 5 平均高出 13.23 分。
单项差距最大的 benchmark：Creative Writing — Claude Opus 5 2,133，Kimi K3 2,082（分差 +50.30）。
本页正文由结构化模型、价格与 benchmark 数据生成，不使用实时 LLM 撰写。