DeepSeek-V4.1-FlashvsClaude Opus 5
在 28 个同模式 benchmark 中，Claude Opus 5 整体领先：DeepSeek-V4.1-Flash 领先 4 项，Claude Opus 5 领先 24 项，持平 0 项，平均分差 -53.37。另有 21 项测试模式不同，仅供参考。
DeepSeek-V4.1-Flash
DeepSeek-AI · 2026-09-10 · 多模态大模型
Claude Opus 5
Anthropic · 2026-07-24 · 推理大模型
DeepSeek-V4.1-Flash4 项(14%)(86%)24 项Claude Opus 5
评测分数
按能力类目分组，每组内按分差大小排列；共 28 项。
仓库修复与多文件工程
Claude Opus 5 领先 2/3| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| APEX-SWE | 51.9014 / 72Max (With Tools) | 63.703 / 72Max (With Tools) | -11.80 |
| NL2Repo-Bench | 65.403 / 20Max (With Tools) | 75.301 / 20Max (With Tools) | -9.90 |
| DeepSWE | 74.208 / 108Max (With Tools) | 73.6514 / 108Max (With Tools) | +0.55 |
图像感知与视觉推理
Claude Opus 5 领先 3/3| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| MMMU-Pro | 7779 / 237Max (No Tools) | 84.7013 / 237Max (No Tools) | -7.70 |
| BabyVision | 89.609 / 17Max (With Tools) | 94.105 / 17Max (With Tools) | -4.50 |
| ZeroBench Main | 4910 / 17Max (With Tools) | 526 / 17Max (With Tools) | -3 |
数学
Claude Opus 5 领先 3/3| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 26.8344 / 77Max (With Tools) | 73.1719 / 77Max (With Tools) | -46.34 |
| Riemann-bench | 2425 / 42Max (With Tools + Internet) | 684 / 42Max (With Tools + Internet) | -44 |
| FrontierMath v2 | 67.3732 / 119Max (With Tools) | 85.6112 / 119Max (With Tools) | -18.25 |
自主开发与终端任务
Claude Opus 5 领先 3/3| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 26.8044 / 110Max (With Tools) | 51.8220 / 110Max (With Tools) | -25.02 |
| Terminal-Bench 3.0 | 304 / 11Max (With Tools) | 43.301 / 11Max (With Tools) | -13.30 |
| Vibe Code Bench 1-100 | 16.3816 / 25Max (With Tools) | 28.533 / 25Max (With Tools) | -12.15 |
抽象归纳与泛化
Claude Opus 5 领先 2/2| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| ARC-AGI-3 | 0.4837 / 66Thinking High (No Tools) | 30.1610 / 66Thinking High (No Tools) | -29.68 |
| ARC-AGI-1 | 88.5093 / 249Thinking High (No Tools) | 97.5013 / 249Thinking High (No Tools) | -9 |
文档与图表理解
Claude Opus 5 领先 2/2| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| GDP.pdf | 12.8074 / 126Max (No Tools) | 21.6035 / 126Max (No Tools) | -8.80 |
| Chartography | 78.9010 / 19Max (With Tools) | 847 / 19Max (With Tools) | -5.10 |
科学计算与科研编程
Claude Opus 5 领先 2/2| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 15.709 / 19Max (With Tools) | 306 / 19Max (With Tools) | -14.30 |
| SciCode | 51.9063 / 136Max (No Tools) | 56.4028 / 136Max (No Tools) | -4.50 |
跨能力综合测试
胶着 2/2| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| LiveBench 2.0 | 78.0412 / 49Max (task-specific tools) | 82.098 / 49Max (task-specific tools) | -4.05 |
| LiveBench | 81.118 / 157Max (No Tools) | 80.0812 / 157Max (No Tools) | +1.03 |
写作与创意表达
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| Creative Writing | 1,54053 / 118Normal (No Tools) | 2,1343 / 118Normal (No Tools) | -593.60 |
办公与文档流程
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| AutomationBench | 54.806 / 148Max (With Tools) | 50.3013 / 148Max (With Tools) | +4.50 |
漏洞利用
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| ExploitGym (budget unspecified) | 15.306 / 10Max (With Tools) | 22.103 / 10Max (With Tools) | -6.80 |
科学知识与推理
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| CritPt | 14.3063 / 207Max (No Tools) | 29.1012 / 207Max (No Tools) | -14.80 |
程序生成与编辑
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| Program Bench | 20.3012 / 18Max (With Tools) | 376 / 18Max (With Tools) | -16.70 |
综合推理
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| GoBench | 1,70717 / 30Max (No Tools) | 2,2567 / 30Max (No Tools) | -549 |
综合知识考试
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| HLE | 63.906 / 247Max (With Tools) | 63.607 / 247Max (With Tools) | +0.30 |
金融
Claude Opus 5 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Claude Opus 5 | 分差 |
|---|---|---|---|
| Tax Agent Bench | 26.6345 / 72Max (With Tools) | 75.062 / 72Max (With Tools) | -48.43 |
测试模式不同的成绩
共 21 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 | DeepSeek-V4.1-Flash | Claude Opus 5 |
|---|---|---|
| IOI (Vals v2) | 40.28Thinking High (With Tools) | 84.33Max (With Tools) |
| BioMysteryBench | 67.78Thinking High (With Tools) | 79.26Max (With Tools) |
| Vals Index | 57.86Thinking High (With Tools) | 67.21Max (With Tools) |
| Vals Index v2.1 | 51.32Thinking High (With Tools) | 63.67Max (task-specific tools) |
| MedCode | 41.17Thinking High (With Tools) | 63.57Max (With Tools) |
| MedScribe | 85.50Thinking High (With Tools) | 90.98Max (With Tools) |
| FrontierCode 1.1 Main | 43.90Thinking High (With Tools) | 53.40Reported best (effort unspecified) |
| Vibe Code Bench v1.1 | 84.74Thinking High (With Tools) | 88.40Max (With Tools) |
| SimpleBench | 66.70Historical report (mode unspecified) | 80.60Thinking High (No Tools) |
| EMB (Excel Modeling) | 57.21Thinking High (With Tools) | 73.56Max (With Tools) |
| Finance Agent v2 | 53.48Thinking High (With Tools) | 58.63Max (With Tools) |
| Harvey's Legal Agent Benchmark | 6.67Thinking High (With Tools) | 6.67Max (With Tools) |
| Legal Research Bench | 41.35Thinking High (With Tools) | 55.29Max (With Tools) |
| Code Migration | 45.62Thinking High (With Tools) | 57.47Max (With Tools) |
| ProofBench v1.1 (Lean 4) | 54Thinking High (With Tools) | 99Max (With Tools) |
| WeirdML v3 | 0.06Thinking High (With Tools) | 0.19Extra-High (With Tools) |
| GPQA Diamond | 90.90Max (No Tools) | 87.88Thinking Low (No Tools) |
| MysteryMechanism | 21.17Thinking High (With Tools) | 37.39Max (With Tools) |
| Public Benefits Bench v1.1 | 64.28Thinking High (With Tools) | 76.93Max (With Tools) |
| SAGE | 47.88Thinking High (With Tools) | 49.43Max (With Tools) |
| APEX-Agents | 39.50Reported best (effort unspecified) | 65.80Max (With Tools) |
规格对比
| 字段 | DeepSeek-V4.1-Flash | Claude Opus 5 |
|---|---|---|
| 发布机构 | DeepSeek-AI | Anthropic |
| 发布时间 | 2026-09-10 | 2026-07-24 |
| 模型类型 | 多模态大模型 | 推理大模型 |
| 架构 | MoE 架构 | 稠密模型 |
| 参数规模 | 5520亿 | 暂无数据 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 384K | 128K |
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 | DeepSeek-V4.1-Flash | Claude Opus 5 |
|---|---|---|
| 文本输入 | ¥1 / 1M tokens | $5 / 1M tokens |
| 文本输出 | ¥4 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | ¥0.02 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $6.25 / 1M tokens |
小结
- DeepSeek-V4.1-Flash在以下类目领先:办公与文档流程 (1/1)、综合知识考试 (1/1)
- Claude Opus 5在以下类目领先:仓库修复与多文件工程 (2/3)、图像感知与视觉推理 (3/3)、数学 (3/3)、自主开发与终端任务 (3/3)、抽象归纳与泛化 (2/2)、文档与图表理解 (2/2)、科学计算与科研编程 (2/2)、写作与创意表达 (1/1)、漏洞利用 (1/1)、科学知识与推理 (1/1)、程序生成与编辑 (1/1)、综合推理 (1/1)、金融 (1/1)
- 胶着类目:跨能力综合测试
28 个同模式 benchmark 上，Claude Opus 5 平均高出 53.37 分。
单项差距最大的 benchmark：Creative Writing — DeepSeek-V4.1-Flash 1,540，Claude Opus 5 2,134（分差 -593.60）。
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。