Claude Opus 5vsGPT-5.6 Sol
在 36 个共同 benchmark 中，Claude Opus 5 整体领先：Claude Opus 5 领先 29 项，GPT-5.6 Sol 领先 7 项，持平 0 项，平均分差 +9.26。
Claude Opus 5
Anthropic · 2026-07-24 · 推理大模型
GPT-5.6 Sol
OpenAI · 2026-06-26 · 推理大模型
Claude Opus 529 项(81%)(19%)7 项GPT-5.6 Sol
评测分数
按能力类目分组，每组内按分差大小排列；共 36 项。
仓库修复与多文件工程
Claude Opus 5 领先 3/4| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| NL2Repo-Bench | 75.301 / 16Max (With Tools) | 56.809 / 16Max (With Tools) | +18.50 |
| SWE-Bench Pro V2 Hard | 981 / 11Extra-High (With Tools) | 82.408 / 11Extra-High (With Tools) | +15.60 |
| SWE-Bench Pro V2 | 99.401 / 10Extra-High (With Tools) | 95.506 / 10Extra-High (With Tools) | +3.90 |
| SWE-Atlas-QnA | 52.704 / 6Reported best (effort unspecified) | 53.503 / 6Reported best (effort unspecified) | -0.80 |
法律
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Legal Research Bench | 55.291 / 42Max (With Tools) | 48.087 / 42Max (With Tools) | +7.21 |
| Harvey Lab-AA | 93.464 / 44Max (With Tools) | 87.1817 / 44Max (With Tools) | +6.28 |
| Harvey's Legal Agent Benchmark | 6.6717 / 43Max (With Tools) | 2.5029 / 43Max (With Tools) | +4.17 |
金融
Claude Opus 5 领先 3/3| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Tax Agent Bench | 75.062 / 20Max (With Tools) | 67.958 / 20Max (With Tools) | +7.11 |
| Finance Agent v2 | 58.638 / 43Max (With Tools) | 53.7621 / 43Max (With Tools) | +4.87 |
| EMB (Excel Modeling) | 73.564 / 42Max (With Tools) | 72.345 / 42Max (With Tools) | +1.22 |
临床工作与医疗决策
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| MedCode | 63.571 / 64Max (With Tools) | 43.9732 / 64Max (With Tools) | +19.60 |
| MedScribe | 90.983 / 66Max (With Tools) | 85.2320 / 66Max (With Tools) | +5.75 |
图像感知与视觉推理
胶着 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| BabyVision | 94.101 / 9Max (With Tools) | 88.904 / 9Max (With Tools) | +5.20 |
| ZeroBench Main | 522 / 6Max (With Tools) | 531 / 6Max (With Tools) | -1 |
多轮记忆与持续上下文
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| EBR-bench | 45.715 / 23Max (With Tools) | 44.766 / 23Max (With Tools) | +0.95 |
| Context Arena | 97.721 / 126Max (No Tools) | 97.632 / 126Max (No Tools) | +0.09 |
客服与业务流程
胶着 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Public Benefits Bench v1.1 | 76.931 / 26Max (With Tools) | 66.5110 / 26Max (With Tools) | +10.42 |
| SAGE | 49.4314 / 64Max (With Tools) | 52.564 / 64Max (With Tools) | -3.13 |
数学
胶着 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| ProofBench v1.1 (Lean 4) | 993 / 28Max (With Tools) | 837 / 28Max (With Tools) | +16 |
| FrontierMath v2 | 85.615 / 58Max (No Tools) | 89.121 / 58Max (No Tools) | -3.51 |
程序生成与编辑
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Program Bench | 376 / 15Max (With Tools) | 238 / 15Max (With Tools) | +14 |
| Vibe Code Bench v1.1 | 88.406 / 61Max (With Tools) | 80.5017 / 61Max (With Tools) | +7.90 |
自主开发与终端任务
Claude Opus 5 领先 2/2| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Terminal-Bench 3.0 | 43.301 / 11Max (With Tools) | 34.402 / 11Max (With Tools) | +8.90 |
| CursorBench 4.0 | 46.605 / 43Max (With Tools) | 41.7010 / 43Max (With Tools) | +4.90 |
专业数据分析
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 53.753 / 29Max (With Tools + Internet) | 47.507 / 29Max (With Tools + Internet) | +6.25 |
写作与创意表达
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Creative Writing | 2,1333 / 110Normal (No Tools) | 1,9729 / 110Normal (No Tools) | +161 |
办公与文档流程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| AutomationBench | 50.306 / 25Max (With Tools) | 45.8011 / 25Max (With Tools) | +4.50 |
文档与图表理解
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Chartography | 843 / 11Max (With Tools) | 79.904 / 11Max (With Tools) | +4.10 |
机器学习工程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| WeirdML v3 | 0.195 / 13Extra-High (With Tools) | 0.157 / 13Extra-High (With Tools) | +0.04 |
漏洞利用
GPT-5.6 Sol 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| ExploitGym (budget unspecified) | 22.102 / 6Max (With Tools) | 33.701 / 6Max (With Tools) | -11.60 |
科学知识与推理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| MysteryMechanism | 37.394 / 13Max (With Tools) | 33.337 / 13Max (With Tools) | +4.06 |
科学计算与科研编程
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 305 / 18Max (With Tools) | 22.406 / 18Max (With Tools) | +7.60 |
算法与竞赛编程
GPT-5.6 Sol 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| IOI (Vals v2) | 84.336 / 26Max (With Tools) | 91.173 / 26Max (With Tools) | -6.84 |
维护、优化与工程管理
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Code Migration | 57.473 / 43Max (With Tools) | 52.927 / 43Max (With Tools) | +4.55 |
综合知识考试
GPT-5.6 Sol 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| HLE-Verified | 54.403 / 6Reported best (effort unspecified) | 54.502 / 6Reported best (effort unspecified) | -0.10 |
跨能力聚合指数
Claude Opus 5 领先 1/1| 评测项 | Claude Opus 5 | GPT-5.6 Sol | 分差 |
|---|---|---|---|
| Vals Index v2.1 | 63.675 / 36Max (task-specific tools) | 58.019 / 36Max (task-specific tools) | +5.66 |
规格对比
| 字段 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|
| 发布机构 | Anthropic | OpenAI |
| 发布时间 | 2026-07-24 | 2026-06-26 |
| 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 | 1.05M |
| 最大输出 | 128K | 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 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|
| 文本输入 | $5 / 1M tokens | $4 / 1M tokens |
| 文本输出 | $25 / 1M tokens | $20 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | $0.4 / 1M tokens |
| 缓存写入 | $6.25 / 1M tokens | $5 / 1M tokens |
小结
- Claude Opus 5在以下类目领先:仓库修复与多文件工程 (3/4)、法律 (3/3)、金融 (3/3)、临床工作与医疗决策 (2/2)、多轮记忆与持续上下文 (2/2)、程序生成与编辑 (2/2)、自主开发与终端任务 (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)、文档与图表理解 (1/1)、机器学习工程 (1/1)、科学知识与推理 (1/1)、科学计算与科研编程 (1/1)、维护、优化与工程管理 (1/1)、跨能力聚合指数 (1/1)
- GPT-5.6 Sol在以下类目领先:漏洞利用 (1/1)、算法与竞赛编程 (1/1)、综合知识考试 (1/1)
- 胶着类目:图像感知与视觉推理、客服与业务流程、数学
36 个共同 benchmark 上，Claude Opus 5 平均高出 9.26 分。
单项差距最大的 benchmark：Creative Writing — Claude Opus 5 2,133，GPT-5.6 Sol 1,972（分差 +161）。
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。