GPT-5.6 SolvsClaude Opus 4.8
在 34 个同模式 benchmark 中，GPT-5.6 Sol 整体领先：GPT-5.6 Sol 领先 23 项，Claude Opus 4.8 领先 11 项，持平 0 项，平均分差 +8.04。另有 9 项测试模式不同，仅供参考。
GPT-5.6 Sol
OpenAI · 2026-06-26 · 推理大模型
Claude Opus 4.8
Anthropic · 2026-05-28 · 推理大模型
GPT-5.6 Sol23 项(68%)(32%)11 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 34 项。
跨应用与工具编排
胶着 4/4| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| MCPMark-Verified | 92.912 / 8Max (With Tools) | 76.386 / 8Max (With Tools) | +16.53 |
| PinchBench v2 | 84.197 / 45Reported best (effort unspecified) | 90.542 / 45Reported best (effort unspecified) | -6.35 |
| Agents' Last Exam | 30.608 / 20Max (With Tools) | 2712 / 20Max (With Tools) | +3.60 |
| MCP-Atlas | 81.8014 / 51Max (With Tools) | 82.2011 / 51Max (With Tools) | -0.40 |
客服与业务流程
Claude Opus 4.8 领先 2/3| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| τ³-Banking | 44.3020 / 168Max (With Tools) | 39.6934 / 168Max (With Tools) | +4.61 |
| SAGE | 52.564 / 65Max (With Tools) | 54.792 / 65Max (With Tools) | -2.23 |
| Public Benefits Bench v1.1 | 66.5110 / 26Max (With Tools) | 68.138 / 26Max (With Tools) | -1.62 |
法律
Claude Opus 4.8 领先 2/3| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Harvey's Legal Agent Benchmark | 2.5035 / 50Max (With Tools) | 9.5814 / 50Max (With Tools) | -7.08 |
| Legal Research Bench | 48.088 / 46Max (With Tools) | 43.7514 / 46Max (With Tools) | +4.33 |
| Harvey Lab-AA | 87.1817 / 44Max (With Tools) | 91.0811 / 44Max (With Tools) | -3.90 |
临床工作与医疗决策
Claude Opus 4.8 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| MedCode | 43.9733 / 66Max (With Tools) | 53.228 / 66Max (With Tools) | -9.25 |
| MedScribe | 85.2320 / 68Max (With Tools) | 85.7517 / 68Max (With Tools) | -0.52 |
仓库修复与多文件工程
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| DeepSWE | 72.6717 / 101Max (With Tools) | 58.9757 / 101Max (With Tools) | +13.70 |
| APEX-SWE | 4818 / 73Max (With Tools) | 43.9031 / 73Max (With Tools) | +4.10 |
多轮记忆与持续上下文
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| EBR-bench | 44.766 / 23Max (With Tools) | 28.5710 / 23Max (With Tools) | +16.19 |
| Context Arena | 97.632 / 126Max (No Tools) | 90.0415 / 126Max (No Tools) | +7.59 |
数学
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 82.9313 / 71Max (With Tools) | 56.1023 / 71Max (With Tools) | +26.83 |
| FrontierMath v2 | 89.126 / 114Max (With Tools) | 8015 / 114Max (With Tools) | +9.12 |
科学知识与推理
胶着 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| CritPt | 32.301 / 204Max (No Tools) | 20.9035 / 204Max (No Tools) | +11.40 |
| GPQA Diamond | 82.83111 / 254Normal (No Tools) | 85.3585 / 254Normal (No Tools) | -2.53 |
科学计算与科研编程
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 22.406 / 18Max (With Tools) | 10.5011 / 18Max (With Tools) | +11.90 |
| SciCode | 57.1022 / 135Max (No Tools) | 54.4045 / 135Max (No Tools) | +2.70 |
自主开发与终端任务
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 37.2728 / 105Max (With Tools) | 23.6440 / 105Max (With Tools) | +13.63 |
| Terminal Bench Hard | 65.901 / 244Max (With Tools) | 58.3010 / 244Max (With Tools) | +7.60 |
金融
胶着 2/2| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| EMB (Excel Modeling) | 72.347 / 46Max (With Tools) | 69.3713 / 46Max (With Tools) | +2.97 |
| Finance Agent v2 | 53.7624 / 46Max (With Tools) | 53.9222 / 46Max (With Tools) | -0.16 |
专业数据分析
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| AA-AnalystAgent | 47.507 / 29Max (With Tools + Internet) | 459 / 29Max (With Tools + Internet) | +2.50 |
写作与创意表达
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Creative Writing | 1,9729 / 112Normal (No Tools) | 1,84018 / 112Normal (No Tools) | +131.80 |
抽象归纳与泛化
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| ARC-AGI-3 | 2.1519 / 60Thinking High (No Tools) | 1.5225 / 60Thinking High (No Tools) | +0.63 |
文档与图表理解
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| GDP.pdf | 27.2012 / 124Max (No Tools) | 22.8028 / 124Max (No Tools) | +4.40 |
程序生成与编辑
Claude Opus 4.8 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 80.5019 / 66Max (With Tools) | 82.7215 / 66Max (With Tools) | -2.22 |
维护、优化与工程管理
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Code Migration | 52.9210 / 48Max (With Tools) | 47.2513 / 48Max (With Tools) | +5.67 |
跨能力综合测试
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| LiveBench | 81.0510 / 157Max (No Tools) | 76.2232 / 157Max (No Tools) | +4.83 |
跨能力聚合指数
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | Claude Opus 4.8 | 分差 |
|---|---|---|---|
| Vals Index v2.1 | 58.019 / 37Max (task-specific tools) | 55.1012 / 37Max (task-specific tools) | +2.91 |
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 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|
| ARC-AGI-1 | 97.50Extra-High (No Tools) | 92.50Max (With Tools) |
| Terminal-Bench 2.1 | 88.80Max (No Tools) | 78.88Thinking High (With Tools) |
| Vals Index | 72.63Extra-High (No Tools) | 60.91Max (With Tools) |
| SimpleBench | 64.80Extra-High (No Tools) | 64.80Normal (No Tools) |
| BrowseComp | 90.40Reported best (effort unspecified) | 84.30Thinking High (With Tools + Internet) |
| HLE | 44.50Max (No Tools) | 57.90Extended (With Tools) |
| WeirdML v2 | 88.76Thinking High (With Tools) | 70.45Normal (With Tools) |
| Text Arena (Coding) | 1,620Extra-High (No Tools) | 1,545Normal (No Tools) |
| SWE-Bench Pro - Public | 64.60Extra-High (With Tools) | 69.20Extended (With Tools) |
规格对比
| 字段 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|
| 发布机构 | OpenAI | Anthropic |
| 发布时间 | 2026-06-26 | 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 | 1.05M | 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 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|
| 文本输入 | $4 / 1M tokens | $5 / 1M tokens |
| 文本输出 | $20 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | $0.4 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | $5 / 1M tokens | $6.25 / 1M tokens |
小结
- GPT-5.6 Sol在以下类目领先:仓库修复与多文件工程 (2/2)、多轮记忆与持续上下文 (2/2)、数学 (2/2)、科学计算与科研编程 (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)、抽象归纳与泛化 (1/1)、文档与图表理解 (1/1)、维护、优化与工程管理 (1/1)、跨能力综合测试 (1/1)、跨能力聚合指数 (1/1)
- Claude Opus 4.8在以下类目领先:客服与业务流程 (2/3)、法律 (2/3)、临床工作与医疗决策 (2/2)、程序生成与编辑 (1/1)
- 胶着类目:跨应用与工具编排、科学知识与推理、金融
34 个同模式 benchmark 上，GPT-5.6 Sol 平均高出 8.04 分。
单项差距最大的 benchmark：Creative Writing — GPT-5.6 Sol 1,972，Claude Opus 4.8 1,840（分差 +131.80）。
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。