Claude Sonnet 5vsClaude Sonnet 4.6
在 21 个同模式 benchmark 中，Claude Sonnet 5 整体领先：Claude Sonnet 5 领先 18 项，Claude Sonnet 4.6 领先 2 项，持平 1 项，平均分差 +5.72。另有 6 项测试模式不同，仅供参考。
Claude Sonnet 5
Anthropic · 2026-06-30 · 多模态大模型
Claude Sonnet 4.6
Anthropic · 2026-02-17 · 聊天大模型
Claude Sonnet 518 项(86%)持平1(10%)2 项Claude Sonnet 4.6
评测分数
按能力类目分组，每组内按分差大小排列；共 21 项。
客服与业务流程
Claude Sonnet 5 领先 3/3| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Public Benefits Bench v1.1 | 66.0311 / 26Max (With Tools) | 62.4515 / 26Max (With Tools) | +3.58 |
| τ³-Banking | 37.3040 / 167Max (With Tools) | 34.4050 / 167Max (With Tools) | +2.90 |
| SAGE | 48.9217 / 64Max (With Tools) | 46.5827 / 64Max (With Tools) | +2.34 |
法律
Claude Sonnet 5 领先 2/3| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Harvey Lab-AA | 90.0714 / 44Max (With Tools) | 85.9719 / 44Max (With Tools) | +4.10 |
| Legal Research Bench | 41.8316 / 46Max (With Tools) | 38.4624 / 46Max (With Tools) | +3.37 |
| Harvey's Legal Agent Benchmark | 527 / 50Max (With Tools) | 527 / 50Max (With Tools) | 持平 |
仓库修复与多文件工程
Claude Sonnet 5 领先 2/2| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| DeepSWE | 48.2374 / 97Thinking High (With Tools) | 29.9390 / 97Thinking High (With Tools) | +18.30 |
| APEX-SWE | 43.6532 / 72Thinking High (With Tools) | 34.5052 / 72Thinking High (With Tools) | +9.15 |
科学知识与推理
Claude Sonnet 5 领先 2/2| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| GPQA Diamond | 80.30131 / 255Max (No Tools) | 78.79141 / 255Max (No Tools) | +1.52 |
| CritPt | 1.10151 / 204Normal (No Tools) | 0.90162 / 204Normal (No Tools) | +0.20 |
金融
Claude Sonnet 5 领先 2/2| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| EMB (Excel Modeling) | 66.3219 / 46Max (With Tools) | 60.1528 / 46Max (With Tools) | +6.17 |
| Finance Agent v2 | 53.9121 / 44Max (With Tools) | 51.0329 / 44Max (With Tools) | +2.88 |
专业数据分析
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 46.258 / 29Max (With Tools + Internet) | 2018 / 29Max (With Tools + Internet) | +26.25 |
写作与创意表达
Claude Sonnet 4.6 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Creative Writing | 1,79424 / 110Normal (No Tools) | 1,81021 / 110Normal (No Tools) | -16.40 |
图像感知与视觉推理
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| MMMU-Pro | 71.90118 / 229Normal (No Tools) | 70.60125 / 229Normal (No Tools) | +1.30 |
文档与图表理解
Claude Sonnet 4.6 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| GDP.pdf | 13.2068 / 122Max (No Tools) | 15.8057 / 122Max (No Tools) | -2.60 |
科学计算与科研编程
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| SciCode | 54.3045 / 134Max (No Tools) | 50.1072 / 134Max (No Tools) | +4.20 |
程序生成与编辑
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 81.3318 / 65Max (With Tools) | 51.4840 / 65Max (With Tools) | +29.85 |
维护、优化与工程管理
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Code Migration | 44.3919 / 48Max (With Tools) | 39.8924 / 48Max (With Tools) | +4.50 |
自主开发与终端任务
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 12.4259 / 102Max (With Tools) | 376 / 102Max (With Tools) | +9.42 |
跨能力聚合指数
Claude Sonnet 5 领先 1/1| 评测项 | Claude Sonnet 5 | Claude Sonnet 4.6 | 分差 |
|---|---|---|---|
| Vals Index | 59.6115 / 42Max (With Tools) | 50.5931 / 42Max (With Tools) | +9.02 |
测试模式不同的成绩
共 6 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 Sonnet 5 | Claude Sonnet 4.6 |
|---|---|---|
| LiveBench | 76.04Extra-High (No Tools) | 75.32Thinking High (No Tools) |
| OSWorld-Verified | 81.20Extra-High (With Tools) | 72.50Thinking (With Tools) |
| BrowseComp | 84.70Thinking (With Tools + Internet) | 74.70Thinking (With Tools) |
| HLE | 57.40Extra-High (With Tools) | 49Thinking (With Tools) |
| Context Arena | 79.53Max (No Tools) | 55.18Normal (No Tools) |
| SWE-bench Verified | 85.20Extra-High (With Tools) | 79.60Thinking (No Tools) |
规格对比
| 字段 | Claude Sonnet 5 | Claude Sonnet 4.6 |
|---|---|---|
| 发布机构 | Anthropic | Anthropic |
| 发布时间 | 2026-06-30 | 2026-02-17 |
| 模型类型 | 多模态大模型 | 聊天大模型 |
| 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 | 8K |
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 Sonnet 5 | Claude Sonnet 4.6 |
|---|---|---|
| 文本输入 | $2 / 1M tokens | $3 / 1M tokens |
| 文本输出 | $10 / 1M tokens | $15 / 1M tokens |
| 缓存读取 | $0.2 / 1M tokens | $0.3 / 1M tokens |
| 缓存写入 | $2.5 / 1M tokens | $3.75 / 1M tokens |
小结
- Claude Sonnet 5在以下类目领先:客服与业务流程 (3/3)、法律 (2/3)、仓库修复与多文件工程 (2/2)、科学知识与推理 (2/2)、金融 (2/2)、专业数据分析 (1/1)、图像感知与视觉推理 (1/1)、科学计算与科研编程 (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)、跨能力聚合指数 (1/1)
- Claude Sonnet 4.6在以下类目领先:写作与创意表达 (1/1)、文档与图表理解 (1/1)
21 个同模式 benchmark 上，Claude Sonnet 5 平均高出 5.72 分。
单项差距最大的 benchmark：Vibe Code Bench v1.1 — Claude Sonnet 5 81.33，Claude Sonnet 4.6 51.48（分差 +29.85）。
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。