GPT-5.5vsOpus 4.7
在 17 个同模式 benchmark 中，GPT-5.5 整体领先：GPT-5.5 领先 11 项，Opus 4.7 领先 6 项，持平 0 项，平均分差 +0.62。另有 26 项测试模式不同，仅供参考。
GPT-5.511 项(65%)(35%)6 项Opus 4.7
评测分数
按能力类目分组，每组内按分差大小排列；共 17 项。
数学
GPT-5.5 领先 4/4| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| MathArena Apex | 80.212 / 17Extra-High (With Tools) | 40.629 / 17Extra-High (With Tools) | +39.59 |
| FrontierMath - Tier 4 | 35.407 / 80Extra-High (No Tools) | 22.9012 / 80Extra-High (No Tools) | +12.50 |
| HMMT Feb 2026 | 98.481 / 11Extra-High (With Tools) | 93.946 / 11Extra-High (With Tools) | +4.54 |
| AIME 2026 | 1001 / 32Extra-High (With Tools) | 95.8310 / 32Extra-High (With Tools) | +4.17 |
抽象归纳与泛化
GPT-5.5 领先 2/2| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 94.5038 / 226Thinking High (No Tools) | 93.5048 / 226Thinking High (No Tools) | +1 |
| ARC-AGI-3 | 0.4335 / 60Thinking High (No Tools) | 0.1847 / 60Thinking High (No Tools) | +0.25 |
科学知识与推理
胶着 2/2| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| GPQA Diamond | 949 / 254Extra-High (No Tools) | 90.1540 / 254Extra-High (No Tools) | +3.85 |
| CritPt | 1.40139 / 204Normal (No Tools) | 5.1098 / 204Normal (No Tools) | -3.70 |
仓库修复与多文件工程
GPT-5.5 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| FrontierSWE v1 | 64.527 / 17Historical report (mode unspecified) | 56.258 / 17Historical report (mode unspecified) | +8.27 |
写作与创意表达
Opus 4.7 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| Creative Writing | 1,84415 / 112Normal (No Tools) | 1,91412 / 112Normal (No Tools) | -70.40 |
图像感知与视觉推理
Opus 4.7 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| MMMU-Pro | 71.40122 / 231Normal (No Tools) | 76.4074 / 231Normal (No Tools) | -5 |
多轮记忆与持续上下文
GPT-5.5 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| Context Arena | 41.9398 / 126Normal (No Tools) | 22.12122 / 126Normal (No Tools) | +19.81 |
常识推理
GPT-5.5 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| SimpleBench | 6920 / 95Normal (No Tools) | 61.7030 / 95Normal (No Tools) | +7.30 |
机器学习工程
Opus 4.7 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| WeirdML v2 | 67.1523 / 53Normal (With Tools) | 76.4013 / 53Normal (With Tools) | -9.25 |
自主开发与终端任务
Opus 4.7 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| Terminal Bench Hard | 49.2024 / 244Normal (With Tools) | 54.5015 / 244Normal (With Tools) | -5.30 |
跨应用与工具编排
Opus 4.7 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| PinchBench v2 | 75.5018 / 45Reported best (effort unspecified) | 76.0115 / 45Reported best (effort unspecified) | -0.51 |
跨能力综合测试
GPT-5.5 领先 1/1| 评测项 | GPT-5.5 | Opus 4.7 | 分差 |
|---|---|---|---|
| LiveBench | 79.9113 / 157Deep Thinking (No Tools) | 76.5331 / 157Deep Thinking (No Tools) | +3.38 |
测试模式不同的成绩
共 26 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-5.5 | Opus 4.7 |
|---|---|---|
| Terminal Bench 2.0 | 82.70Thinking High (With Tools) | 69.40Extended (With Tools) |
| Terminal-Bench 2.1 | 83.10Extra-High (With Tools) | 69.70Thinking High (With Tools) |
| Vals Index | 57.41Extra-High (With Tools) | 56.11Max (With Tools) |
| MedCode | 49.10Extra-High (With Tools) | 54.86Max (With Tools) |
| MedScribe | 86.87Extra-High (With Tools) | 82.95Max (With Tools) |
| Vibe Code Bench v1.1 | 69.85Extra-High (With Tools) | 71Max (With Tools) |
| AA-AnalystAgent | 7.50Normal (With Tools + Internet) | 43.75Max (With Tools + Internet) |
| OSWorld-Verified | 78.70Thinking High (With Tools) | 78Extended (With Tools) |
| BrowseComp | 84.40Thinking High (With Tools + Internet) | 79.30Extended (With Tools) |
| EMB (Excel Modeling) | 64.54Extra-High (With Tools) | 63.73Max (With Tools) |
| Finance Agent v2 | 51.76Extra-High (With Tools) | 51.51Max (With Tools) |
| HLE | 52.20Thinking High (With Tools) | 54.70Extended (With Tools) |
| Harvey's Legal Agent Benchmark | 3.75Extra-High (With Tools) | 6.67Max (With Tools) |
| Legal Research Bench | 40.38Extra-High (With Tools) | 38.46Max (With Tools) |
| Code Migration | 45.16Extra-High (With Tools) | 43.88Max (With Tools) |
| GSO | 40.20Extra-High (With Tools) | 44.12Normal (With Tools) |
| FrontierMath | 51.70Thinking High (With Tools) | 43.80Extra-High (No Tools) |
| FrontierMath Tier 4 v2 | 72.50Extra-High (With Tools) | 31.71Max (With Tools) |
| FrontierMath v2 | 85.26Extra-High (With Tools) | 70.18Max (With Tools) |
| EBR-bench | 34.29Extra-High (With Tools) | 19.05Max (With Tools) |
| Text Arena (Coding) | 1,505Extra-High (No Tools) | 1,562Normal (No Tools) |
| APEX-SWE | 37Extra-High (With Tools) | 47.60Max (With Tools) |
| SWE-Bench Pro - Public | 58.60Thinking High (With Tools) | 64.30Extended (With Tools) |
| SAGE | 51.53Extra-High (With Tools) | 56.10Max (With Tools) |
| τ³-Banking | 14.80Normal (With Tools) | 40.21Max (With Tools) |
| MCP-Atlas | 75.30Extra-High (With Tools) | 79.10Deep Thinking (With Tools) |
规格对比
| 字段 | GPT-5.5 | Opus 4.7 |
|---|---|---|
| 发布机构 | OpenAI | Anthropic |
| 发布时间 | 2026-04-23 | 2026-04-16 |
| 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 | 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 | GPT-5.5 | Opus 4.7 |
|---|---|---|
| 文本输入 | $5 / 1M tokens | $5 / 1M tokens |
| 文本输出 | $30 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $6.25 / 1M tokens |
小结
- GPT-5.5在以下类目领先:数学 (4/4)、抽象归纳与泛化 (2/2)、仓库修复与多文件工程 (1/1)、多轮记忆与持续上下文 (1/1)、常识推理 (1/1)、跨能力综合测试 (1/1)
- Opus 4.7在以下类目领先: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)、自主开发与终端任务 (1/1)、跨应用与工具编排 (1/1)
- 胶着类目:科学知识与推理
17 个同模式 benchmark 上，GPT-5.5 平均高出 0.62 分。
单项差距最大的 benchmark：Creative Writing — GPT-5.5 1,844，Opus 4.7 1,914（分差 -70.40）。
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。