Kimi K3vsClaude Opus 4.8
在 31 个同模式 benchmark 中，Kimi K3 整体领先：Kimi K3 领先 17 项，Claude Opus 4.8 领先 14 项，持平 0 项，平均分差 +6.03。另有 8 项测试模式不同，仅供参考。
Kimi K3
Moonshot AI · 2026-07-16 · 推理大模型
Claude Opus 4.8
Anthropic · 2026-05-28 · 推理大模型
Kimi K317 项(55%)(45%)14 项Claude Opus 4.8
评测分数
按能力类目分组，每组内按分差大小排列；共 31 项。
跨应用与工具编排
Kimi K3 领先 4/4| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| MCPMark-Verified | 94.501 / 8Max (With Tools) | 76.386 / 8Max (With Tools) | +18.12 |
| Agents' Last Exam | 28.3010 / 20Max (With Tools) | 2712 / 20Max (With Tools) | +1.30 |
| Toolathlon-Verified | 76.506 / 36Max (With Tools) | 76.207 / 36Max (With Tools) | +0.30 |
| MCP-Atlas | 82.3010 / 53Max (With Tools) | 82.2011 / 53Max (With Tools) | +0.10 |
客服与业务流程
Claude Opus 4.8 领先 2/3| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| τ³-Banking | 37.1142 / 171Max (With Tools) | 39.6934 / 171Max (With Tools) | -2.58 |
| SAGE | 54.263 / 65Max (With Tools) | 54.792 / 65Max (With Tools) | -0.53 |
| Public Benefits Bench v1.1 | 68.277 / 26Max (With Tools) | 68.138 / 26Max (With Tools) | +0.14 |
法律
Kimi K3 领先 3/3| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Harvey Lab-AA | 94.642 / 44Max (With Tools) | 91.0811 / 44Max (With Tools) | +3.56 |
| Harvey's Legal Agent Benchmark | 12.928 / 50Max (With Tools) | 9.5814 / 50Max (With Tools) | +3.34 |
| Legal Research Bench | 44.2013 / 46Max (With Tools) | 43.7514 / 46Max (With Tools) | +0.45 |
临床工作与医疗决策
胶着 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| MedCode | 48.8820 / 66Max (With Tools) | 53.228 / 66Max (With Tools) | -4.34 |
| MedScribe | 87.9610 / 68Max (With Tools) | 85.7517 / 68Max (With Tools) | +2.21 |
仓库修复与多文件工程
Kimi K3 领先 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| DeepSWE | 68.5132 / 103Max (With Tools) | 58.9757 / 103Max (With Tools) | +9.54 |
| APEX-SWE | 4818 / 73Max (With Tools) | 43.9031 / 73Max (With Tools) | +4.10 |
数学
Claude Opus 4.8 领先 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 39.0230 / 71Max (With Tools) | 56.1023 / 71Max (With Tools) | -17.07 |
| FrontierMath v2 | 72.1822 / 114Max (With Tools) | 8015 / 114Max (With Tools) | -7.82 |
科学知识与推理
胶着 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| CritPt | 23.4030 / 206Max (No Tools) | 20.9035 / 206Max (No Tools) | +2.50 |
| GPQA Diamond | 91.9224 / 258Thinking High (No Tools) | 93.6013 / 258Thinking High (No Tools) | -1.68 |
科学计算与科研编程
胶着 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| SciCode | 59.508 / 135Max (No Tools) | 54.4045 / 135Max (No Tools) | +5.10 |
| Terminal-Bench-Science 0.1 | 7.1014 / 18Max (With Tools) | 10.5011 / 18Max (With Tools) | -3.40 |
金融
胶着 2/2| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| EMB (Excel Modeling) | 66.4018 / 46Max (With Tools) | 69.3713 / 46Max (With Tools) | -2.97 |
| Finance Agent v2 | 54.4021 / 46Max (With Tools) | 53.9222 / 46Max (With Tools) | +0.48 |
专业数据分析
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 38.7514 / 29Max (With Tools + Internet) | 459 / 29Max (With Tools + Internet) | -6.25 |
写作与创意表达
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Creative Writing | 2,0825 / 112Normal (No Tools) | 1,84018 / 112Normal (No Tools) | +242.50 |
多轮记忆与持续上下文
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Context Arena | 71.7559 / 126Max (No Tools) | 90.0415 / 126Max (No Tools) | -18.29 |
抽象归纳与泛化
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 94.5038 / 226Max (With Tools) | 92.5053 / 226Max (With Tools) | +2 |
文档与图表理解
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| GDP.pdf | 2232 / 124Max (No Tools) | 22.8028 / 124Max (No Tools) | -0.80 |
程序生成与编辑
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 84.9612 / 66Max (With Tools) | 82.7215 / 66Max (With Tools) | +2.24 |
维护、优化与工程管理
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Code Migration | 16.1039 / 48Max (With Tools) | 47.2513 / 48Max (With Tools) | -31.15 |
自主开发与终端任务
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 12.6059 / 105Max (With Tools) | 23.6440 / 105Max (With Tools) | -11.04 |
跨能力聚合指数
Claude Opus 4.8 领先 1/1| 评测项 | Kimi K3 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vals Index | 57.8120 / 42Max (With Tools) | 60.9111 / 42Max (With Tools) | -3.10 |
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 | Kimi K3 | Claude Opus 4.8 |
|---|---|---|
| Vals Index v2.1 | 50.70Reported best (effort unspecified) | 55.10Max (task-specific tools) |
| SimpleBench | 60.70Max (No Tools) | 64.80Normal (No Tools) |
| LiveBench | 79.19Reported best (effort unspecified) | 78.93Deep Thinking (No Tools) |
| OSWorld-Verified | 84.80Max (With Tools) | 83.40Extended (With Tools) |
| BrowseComp | 91.20Max (With Tools + Internet) | 84.30Thinking High (With Tools + Internet) |
| HLE | 59.80Max (With Tools) | 57.90Extended (With Tools) |
| MathArena Apex | 65.60Max (No Tools) | 81.25Max (With Tools) |
| Text Arena (Coding) | 1,682Max (No Tools) | 1,545Normal (No Tools) |
规格对比
| 字段 | Kimi K3 | Claude Opus 4.8 |
|---|---|---|
| 发布机构 | Moonshot AI | Anthropic |
| 发布时间 | 2026-07-16 | 2026-05-28 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | MoE 架构 | 稠密模型 |
| 参数规模 | 2.8万亿 | 暂无数据 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 1M | 125K |
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 | Kimi K3 | Claude Opus 4.8 |
|---|---|---|
| 文本输入 | ¥20 / 1M tokens | $5 / 1M tokens |
| 文本输出 | ¥100 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | ¥2 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $6.25 / 1M tokens |
小结
- Kimi K3在以下类目领先:跨应用与工具编排 (4/4)、法律 (3/3)、仓库修复与多文件工程 (2/2)、写作与创意表达 (1/1)、抽象归纳与泛化 (1/1)、程序生成与编辑 (1/1)
- Claude Opus 4.8在以下类目领先:客服与业务流程 (2/3)、数学 (2/2)、专业数据分析 (1/1)、多轮记忆与持续上下文 (1/1)、文档与图表理解 (1/1)、维护、优化与工程管理 (1/1)、自主开发与终端任务 (1/1)、跨能力聚合指数 (1/1)
- 胶着类目:临床工作与医疗决策、科学知识与推理、科学计算与科研编程、金融
31 个同模式 benchmark 上，Kimi K3 平均高出 6.03 分。
单项差距最大的 benchmark：Creative Writing — Kimi K3 2,082，Claude Opus 4.8 1,840（分差 +242.50）。
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。