Opus 4.7vsClaude Opus 4.6
在 27 个同模式 benchmark 中，Opus 4.7 整体领先：Opus 4.7 领先 23 项，Claude Opus 4.6 领先 4 项，持平 0 项，平均分差 +7.22。另有 8 项测试模式不同，仅供参考。
Opus 4.7
Anthropic · 2026-04-16 · 推理大模型
Claude Opus 4.6
Anthropic · 2026-02-05 · 推理大模型
Opus 4.723 项(85%)(15%)4 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 27 项。
仓库修复与多文件工程
Opus 4.7 领先 3/3| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| FrontierSWE v1 | 56.258 / 17Historical report (mode unspecified) | 48.909 / 17Historical report (mode unspecified) | +7.35 |
| SWE-bench Verified | 87.606 / 117Extended (With Tools) | 80.8410 / 117Extended (With Tools) | +6.76 |
| APEX-SWE | 47.6020 / 73Max (With Tools) | 41.8034 / 73Max (With Tools) | +5.80 |
临床工作与医疗决策
胶着 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| MedCode | 54.865 / 66Max (With Tools) | 49.1318 / 66Max (With Tools) | +5.73 |
| MedScribe | 82.9534 / 68Max (With Tools) | 86.1316 / 68Max (With Tools) | -3.18 |
多轮记忆与持续上下文
胶着 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Context Arena | 22.12122 / 126Normal (No Tools) | 60.3376 / 126Normal (No Tools) | -38.21 |
| EBR-bench | 19.0513 / 23Max (With Tools) | 12.7016 / 23Max (With Tools) | +6.35 |
客服与业务流程
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| τ³-Banking | 40.2131 / 168Max (With Tools) | 27.3270 / 168Max (With Tools) | +12.89 |
| SAGE | 56.101 / 65Max (With Tools) | 51.588 / 65Max (With Tools) | +4.52 |
数学
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 31.7134 / 71Max (With Tools) | 26.8340 / 71Max (With Tools) | +4.88 |
| FrontierMath v2 | 70.1824 / 114Max (With Tools) | 65.9628 / 114Max (With Tools) | +4.21 |
科学知识与推理
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| GPQA Diamond | 94.207 / 253Extended (No Tools) | 91.3129 / 253Extended (No Tools) | +2.89 |
| CritPt | 5.1098 / 204Normal (No Tools) | 2.80124 / 204Normal (No Tools) | +2.30 |
自主开发与终端任务
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Terminal Bench Hard | 54.5015 / 244Normal (With Tools) | 48.5025 / 244Normal (With Tools) | +6 |
| Terminal Bench 2.0 | 69.406 / 51Extended (With Tools) | 65.4011 / 51Extended (With Tools) | +4 |
跨应用与工具编排
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| PinchBench v2 | 76.0115 / 45Reported best (effort unspecified) | 69.9026 / 45Reported best (effort unspecified) | +6.11 |
| MCP-Atlas | 79.1014 / 50Deep Thinking (With Tools) | 76.8019 / 50Deep Thinking (With Tools) | +2.30 |
写作与创意表达
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Creative Writing | 1,91412 / 112Normal (No Tools) | 1,80922 / 112Normal (No Tools) | +105.20 |
图像感知与视觉推理
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| MMMU-Pro | 76.4074 / 231Normal (No Tools) | 72.50117 / 231Normal (No Tools) | +3.90 |
常识推理
Claude Opus 4.6 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| SimpleBench | 61.7030 / 95Normal (No Tools) | 67.6021 / 95Normal (No Tools) | -5.90 |
抽象归纳与泛化
Claude Opus 4.6 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 93.5048 / 221Thinking High (No Tools) | 9444 / 221Thinking High (No Tools) | -0.50 |
机器学习工程
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| WeirdML v2 | 76.4013 / 53Normal (With Tools) | 65.9024 / 53Normal (With Tools) | +10.50 |
桌面工作流
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| OSWorld-Verified | 7813 / 28Extended (With Tools) | 72.7018 / 28Extended (With Tools) | +5.30 |
程序生成与编辑
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 7129 / 66Max (With Tools) | 53.5038 / 66Max (With Tools) | +17.50 |
维护、优化与工程管理
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| GSO | 44.121 / 22Normal (With Tools) | 33.335 / 22Normal (With Tools) | +10.79 |
综合偏好评测
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Text Arena (Coding) | 1,5627 / 36Normal (No Tools) | 1,5558 / 36Normal (No Tools) | +7.04 |
跨能力综合测试
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| LiveBench | 74.8945 / 157Thinking High (No Tools) | 74.5248 / 157Thinking High (No Tools) | +0.37 |
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 | Opus 4.7 | Claude Opus 4.6 |
|---|---|---|
| ARC-AGI-3 | 0.18Thinking High (No Tools) | 0.51Max (No Tools) |
| BrowseComp | 79.30Extended (With Tools) | 84Thinking (With Tools + Internet) |
| HLE | 54.70Extended (With Tools) | 19Normal (No Tools) |
| AIME 2026 | 95.83Extra-High (With Tools) | 96.67Thinking High (With Tools) |
| FrontierMath | 43.80Extra-High (No Tools) | 40.70Max (No Tools) |
| FrontierMath - Tier 4 | 22.90Extra-High (No Tools) | 22.90Max (No Tools) |
| HMMT Feb 2026 | 93.94Extra-High (With Tools) | 96.21Thinking High (With Tools) |
| MathArena Apex | 40.62Extra-High (With Tools) | 34.45Thinking High (With Tools) |
规格对比
| 字段 | Opus 4.7 | Claude Opus 4.6 |
|---|---|---|
| 发布机构 | Anthropic | Anthropic |
| 发布时间 | 2026-04-16 | 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 | 1000K | 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 | Opus 4.7 | 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
- Opus 4.7在以下类目领先:仓库修复与多文件工程 (3/3)、客服与业务流程 (2/2)、数学 (2/2)、科学知识与推理 (2/2)、自主开发与终端任务 (2/2)、跨应用与工具编排 (2/2)、写作与创意表达 (1/1)、图像感知与视觉推理 (1/1)、机器学习工程 (1/1)、桌面工作流 (1/1)、程序生成与编辑 (1/1)、维护、优化与工程管理 (1/1)、综合偏好评测 (1/1)、跨能力综合测试 (1/1)
- Claude Opus 4.6在以下类目领先:常识推理 (1/1)、抽象归纳与泛化 (1/1)
- 胶着类目:临床工作与医疗决策、多轮记忆与持续上下文
27 个同模式 benchmark 上，Opus 4.7 平均高出 7.22 分。
单项差距最大的 benchmark：Creative Writing — Opus 4.7 1,914，Claude Opus 4.6 1,809（分差 +105.20）。
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。