Claude Opus 5vsClaude Opus 4.8
在 38 个同模式 benchmark 中，Claude Opus 5 整体领先：Claude Opus 5 领先 34 项，Claude Opus 4.8 领先 4 项，持平 0 项，平均分差 +17.62。另有 9 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 4.8
Anthropic · 2026-05-28 · 推理大模型
Claude Opus 534 项(89%)(11%)4 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 38 项。
客服与业务流程
Claude Opus 5 领先 2/3| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| τ³-Banking | 48.716 / 170Max (With Tools) | 39.6934 / 170Max (With Tools) | +9.02 |
| Public Benefits Bench v1.1 | 76.931 / 50Max (With Tools) | 68.1313 / 50Max (With Tools) | +8.80 |
| SAGE | 49.4321 / 91Max (With Tools) | 54.793 / 91Max (With Tools) | -5.36 |
数学
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Riemann-bench | 684 / 42Max (With Tools + Internet) | 47.2012 / 42Max (With Tools + Internet) | +20.80 |
| FrontierMath Tier 4 v2 | 73.1719 / 77Max (With Tools) | 56.1025 / 77Max (With Tools) | +17.07 |
| FrontierMath v2 | 85.6112 / 120Max (With Tools) | 8017 / 120Max (With Tools) | +5.61 |
法律
Claude Opus 5 领先 2/3| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Legal Research Bench | 55.291 / 84Max (With Tools) | 43.7522 / 84Max (With Tools) | +11.54 |
| Harvey's Legal Agent Benchmark | 6.6721 / 51Max (With Tools) | 9.5814 / 51Max (With Tools) | -2.91 |
| Harvey Lab-AA | 93.464 / 44Max (With Tools) | 91.0811 / 44Max (With Tools) | +2.38 |
金融
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Tax Agent Bench | 45.5524 / 71Unspecified Reasoning (With Tools) | 3040 / 71Unspecified Reasoning (With Tools) | +15.55 |
| Finance Agent v2 | 58.6310 / 76Max (With Tools) | 53.9229 / 76Max (With Tools) | +4.71 |
| EMB (Excel Modeling) | 73.566 / 77Max (With Tools) | 69.3715 / 77Max (With Tools) | +4.19 |
临床工作与医疗决策
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| MedCode | 63.571 / 107Max (With Tools) | 53.2210 / 107Max (With Tools) | +10.35 |
| MedScribe | 90.984 / 109Max (With Tools) | 85.7526 / 109Max (With Tools) | +5.23 |
仓库修复与多文件工程
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| APEX-SWE | 63.703 / 72Max (With Tools) | 43.9031 / 72Max (With Tools) | +19.80 |
| DeepSWE | 73.6512 / 106Max (With Tools) | 58.9760 / 106Max (With Tools) | +14.68 |
多轮记忆与持续上下文
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| EBR-bench | 45.715 / 23Max (With Tools) | 28.5710 / 23Max (With Tools) | +17.14 |
| Context Arena | 97.721 / 154Max (No Tools) | 90.0415 / 154Max (No Tools) | +7.68 |
抽象归纳与泛化
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| ARC-AGI-3 | 30.1610 / 66Thinking High (No Tools) | 1.5228 / 66Thinking High (No Tools) | +28.64 |
| ARC-AGI-1 | 97.5013 / 249Max (With Tools) | 92.5059 / 249Max (With Tools) | +5 |
科学知识与推理
胶着 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| CritPt | 29.1012 / 208Max (No Tools) | 20.9037 / 208Max (No Tools) | +8.20 |
| GPQA Diamond | 87.8865 / 250Thinking Low (No Tools) | 88.3860 / 250Thinking Low (No Tools) | -0.50 |
科学计算与科研编程
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 305 / 18Max (With Tools) | 10.5011 / 18Max (With Tools) | +19.50 |
| SciCode | 56.4028 / 137Max (No Tools) | 54.4046 / 137Max (No Tools) | +2 |
跨应用与工具编排
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| APEX-Agents | 65.806 / 61Max (With Tools) | 48.9031 / 61Max (With Tools) | +16.90 |
| Agents' Last Exam | 32.205 / 21Max (With Tools) | 2713 / 21Max (With Tools) | +5.20 |
跨能力综合测试
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| LiveBench 2.0 | 82.098 / 49Max (task-specific tools) | 76.7713 / 49Max (task-specific tools) | +5.32 |
| LiveBench | 80.0812 / 157Max (No Tools) | 76.2232 / 157Max (No Tools) | +3.86 |
跨能力聚合指数
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vals Index v2.1 | 63.675 / 38Max (task-specific tools) | 55.1013 / 38Max (task-specific tools) | +8.57 |
| Vals Index | 67.214 / 42Max (With Tools) | 60.9111 / 42Max (With Tools) | +6.30 |
专业数据分析
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 53.753 / 29Max (With Tools + Internet) | 459 / 29Max (With Tools + Internet) | +8.75 |
写作与创意表达
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Creative Writing | 2,1333 / 114Normal (No Tools) | 1,84018 / 114Normal (No Tools) | +292.80 |
办公与文档流程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| AutomationBench | 50.3013 / 148Max (With Tools) | 16.4472 / 148Max (With Tools) | +33.86 |
文档与图表理解
Claude Opus 4.8 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| GDP.pdf | 21.6034 / 126Max (No Tools) | 22.8028 / 126Max (No Tools) | -1.20 |
机器学习工程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| WeirdML v2 | 86.344 / 52Normal (With Tools) | 70.4518 / 52Normal (With Tools) | +15.89 |
程序生成与编辑
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 88.4010 / 114Max (With Tools) | 82.7222 / 114Max (With Tools) | +5.68 |
维护、优化与工程管理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Code Migration | 57.478 / 79Max (With Tools) | 47.2516 / 79Max (With Tools) | +10.22 |
自主开发与终端任务
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 51.8217 / 108Max (With Tools) | 23.6443 / 108Max (With Tools) | +28.18 |
测试模式不同的成绩
共 9 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 4.8 |
|---|---|---|
| FrontierCode 1.1 Main | 53.40Reported best (effort unspecified) | 46.50Max (With Tools) |
| SimpleBench | 80.60Thinking High (No Tools) | 64.80Normal (No Tools) |
| BrowseComp | 90.80Max (With Tools + Internet) | 84.30Thinking High (With Tools + Internet) |
| HLE | 63.60Max (With Tools) | 57.90Extended (With Tools) |
| PostTrain Bench | 35Reported best (effort unspecified) | 33.80Thinking High (With Tools) |
| Text Arena (Coding) | 1,712Max (No Tools) | 1,545Normal (No Tools) |
| SWE-Bench Pro - Public | 79.20Max (With Tools) | 69.20Extended (With Tools) |
| SWE-bench Verified | 96Max (With Tools) | 88.60Extended (With Tools) |
| MCP-Atlas | 85.80Extra-High (With Tools) | 82.20Deep Thinking (With Tools) |
规格对比
| 字段 | Claude Opus 5 | Claude Opus 4.8 |
|---|---|---|
| 发布机构 | Anthropic | Anthropic |
| 发布时间 | 2026-07-24 | 2026-05-28 |
| 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 | 1M |
| 最大输出 | 128K | 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 | Claude Opus 5 | Claude Opus 4.8 |
|---|---|---|
| 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:客服与业务流程 (2/3)、数学 (3/3)、法律 (2/3)、金融 (3/3)、临床工作与医疗决策 (2/2)、仓库修复与多文件工程 (2/2)、多轮记忆与持续上下文 (2/2)、抽象归纳与泛化 (2/2)、科学计算与科研编程 (2/2)、跨应用与工具编排 (2/2)、跨能力综合测试 (2/2)、跨能力聚合指数 (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)、机器学习工程 (1/1)、程序生成与编辑 (1/1)、维护、优化与工程管理 (1/1)、自主开发与终端任务 (1/1)
- Claude Opus 4.8在以下类目领先:文档与图表理解 (1/1)
- 胶着类目:科学知识与推理
38 个同模式 benchmark 上，Claude Opus 5 平均高出 17.62 分。
单项差距最大的 benchmark：Creative Writing — Claude Opus 5 2,133，Claude Opus 4.8 1,840（分差 +292.80）。
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。