Claude Opus 5vsClaude Opus 4.6
在 17 个同模式 benchmark 中，Claude Opus 5 整体领先：Claude Opus 5 领先 16 项，Claude Opus 4.6 领先 1 项，持平 0 项，平均分差 +38.22。另有 11 项测试模式不同，仅供参考。
Claude Opus 5
Anthropic · 2026-07-24 · 推理大模型
Claude Opus 4.6
Anthropic · 2026-02-05 · 推理大模型
Claude Opus 516 项(94%)(6%)1 项Claude Opus 4.6
评测分数
按能力类目分组，每组内按分差大小排列；共 17 项。
数学
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 73.1719 / 77Max (With Tools) | 26.8344 / 77Max (With Tools) | +46.34 |
| Riemann-bench | 684 / 42Max (With Tools + Internet) | 27.2024 / 42Max (With Tools + Internet) | +40.80 |
| FrontierMath v2 | 85.6112 / 120Max (With Tools) | 65.9633 / 120Max (With Tools) | +19.65 |
临床工作与医疗决策
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| MedCode | 63.571 / 107Max (With Tools) | 49.1325 / 107Max (With Tools) | +14.44 |
| MedScribe | 90.984 / 109Max (With Tools) | 86.1324 / 109Max (With Tools) | +4.85 |
客服与业务流程
胶着 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| τ³-Banking | 48.716 / 170Max (With Tools) | 27.3271 / 170Max (With Tools) | +21.39 |
| SAGE | 49.4321 / 91Max (With Tools) | 51.5812 / 91Max (With Tools) | -2.15 |
仓库修复与多文件工程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| APEX-SWE | 63.703 / 72Max (With Tools) | 41.8034 / 72Max (With Tools) | +21.90 |
写作与创意表达
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Creative Writing | 2,1333 / 114Normal (No Tools) | 1,80922 / 114Normal (No Tools) | +323.50 |
图像感知与视觉推理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| MMMU-Pro | 84.708 / 233Max (No Tools) | 75.4089 / 233Max (No Tools) | +9.30 |
多轮记忆与持续上下文
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| EBR-bench | 45.715 / 23Max (With Tools) | 12.7016 / 23Max (With Tools) | +33.01 |
抽象归纳与泛化
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 97.5013 / 249Thinking High (No Tools) | 9449 / 249Thinking High (No Tools) | +3.50 |
机器学习工程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| WeirdML v2 | 86.344 / 52Normal (With Tools) | 65.9023 / 52Normal (With Tools) | +20.44 |
科学知识与推理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| CritPt | 29.1012 / 208Max (No Tools) | 12.6069 / 208Max (No Tools) | +16.50 |
程序生成与编辑
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 88.4010 / 114Max (With Tools) | 53.5054 / 114Max (With Tools) | +34.90 |
综合知识考试
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| HLE | 56.3024 / 244Max (No Tools) | 34.44116 / 244Max (No Tools) | +21.86 |
跨应用与工具编排
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| APEX-Agents | 65.806 / 61Max (With Tools) | 46.3038 / 61Max (With Tools) | +19.50 |
测试模式不同的成绩
共 11 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 | Claude Opus 4.6 |
|---|---|---|
| ARC-AGI-3 | 30.16Thinking High (No Tools) | 0.51Max (No Tools) |
| FrontierCode 1.1 Main | 53.40Reported best (effort unspecified) | 26.60Thinking High (With Tools) |
| SimpleBench | 80.60Thinking High (No Tools) | 67.60Normal (No Tools) |
| LiveBench | 80.08Max (No Tools) | 74.52Thinking High (No Tools) |
| BrowseComp | 90.80Max (With Tools + Internet) | 84Thinking (With Tools + Internet) |
| Context Arena | 97.72Max (No Tools) | 60.33Normal (No Tools) |
| Text Arena (Coding) | 1,712Max (No Tools) | 1,555Normal (No Tools) |
| SWE-bench Multilingual | 89.50Max (With Tools) | 77.83Extended (With Tools) |
| SWE-bench Verified | 96Max (With Tools) | 80.84Extended (With Tools) |
| GPQA Diamond | 87.88Thinking Low (No Tools) | 91.31Extended (No Tools) |
| MCP-Atlas | 85.80Extra-High (With Tools) | 76.80Deep Thinking (With Tools) |
规格对比
| 字段 | Claude Opus 5 | Claude Opus 4.6 |
|---|---|---|
| 发布机构 | Anthropic | Anthropic |
| 发布时间 | 2026-07-24 | 2026-02-05 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | 稠密模型 | 稠密模型 |
| 参数规模 | 暂无数据 | 暂无数据 |
| 上下文长度 | 1M | 1000K |
| 最大输出 | 128K | 64K |
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 | Claude Opus 4.6 |
|---|---|---|
| 文本输入 | $5 / 1M tokens | $5 / 1M tokens |
| 文本输出 | $25 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | $6.25 / 1M tokens | $6.25 / 1M tokens |
小结
- Claude Opus 5在以下类目领先:数学 (3/3)、临床工作与医疗决策 (2/2)、仓库修复与多文件工程 (1/1)、写作与创意表达 (1/1)、图像感知与视觉推理 (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)、科学知识与推理 (1/1)、程序生成与编辑 (1/1)、综合知识考试 (1/1)、跨应用与工具编排 (1/1)
- 胶着类目:客服与业务流程
17 个同模式 benchmark 上，Claude Opus 5 平均高出 38.22 分。
单项差距最大的 benchmark：Creative Writing — Claude Opus 5 2,133，Claude Opus 4.6 1,809（分差 +323.50）。
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。