Opus 4.7vsGPT-5.4
在 18 个同模式 benchmark 中，Opus 4.7 整体领先：Opus 4.7 领先 10 项，GPT-5.4 领先 8 项，持平 0 项，平均分差 +6.58。另有 17 项测试模式不同，仅供参考。
Opus 4.710 项(56%)(44%)8 项GPT-5.4
评测分数
按能力类目分组，每组内按分差大小排列；共 18 项。
数学
GPT-5.4 领先 5/5| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| MathArena Apex | 40.629 / 17Extra-High (With Tools) | 54.178 / 17Extra-High (With Tools) | -13.55 |
| FrontierMath - Tier 4 | 22.9012 / 80Extra-High (No Tools) | 27.1011 / 80Extra-High (No Tools) | -4.20 |
| FrontierMath | 43.806 / 61Extra-High (No Tools) | 47.605 / 61Extra-High (No Tools) | -3.80 |
| HMMT Feb 2026 | 93.946 / 11Extra-High (With Tools) | 97.732 / 11Extra-High (With Tools) | -3.79 |
| AIME 2026 | 95.8310 / 32Extra-High (With Tools) | 99.174 / 32Extra-High (With Tools) | -3.34 |
仓库修复与多文件工程
Opus 4.7 领先 2/2| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| FrontierSWE v1 | 56.258 / 17Historical report (mode unspecified) | 45.9610 / 17Historical report (mode unspecified) | +10.29 |
| APEX-SWE | 41.3436 / 73Thinking High (With Tools) | 36.5047 / 73Thinking High (With Tools) | +4.84 |
抽象归纳与泛化
胶着 2/2| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 9166 / 221Thinking Medium (No Tools) | 86.17100 / 221Thinking Medium (No Tools) | +4.83 |
| ARC-AGI-3 | 0.1842 / 55Thinking High (No Tools) | 0.2140 / 55Thinking High (No Tools) | -0.03 |
写作与创意表达
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| Creative Writing | 1,91412 / 112Normal (No Tools) | 1,84019 / 112Normal (No Tools) | +74.60 |
图像感知与视觉推理
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| MMMU-Pro | 76.4074 / 231Normal (No Tools) | 70.60126 / 231Normal (No Tools) | +5.80 |
多轮记忆与持续上下文
GPT-5.4 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| Context Arena | 22.12122 / 126Normal (No Tools) | 32.79112 / 126Normal (No Tools) | -10.67 |
机器学习工程
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| WeirdML v2 | 76.4013 / 53Normal (With Tools) | 57.4432 / 53Normal (With Tools) | +18.96 |
科学知识与推理
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| CritPt | 5.1098 / 204Normal (No Tools) | 0.60172 / 204Normal (No Tools) | +4.50 |
维护、优化与工程管理
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| GSO | 44.102 / 22Thinking High (With Tools) | 25.499 / 22Thinking High (With Tools) | +18.61 |
自主开发与终端任务
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| Terminal Bench Hard | 54.5015 / 244Normal (With Tools) | 37.9059 / 244Normal (With Tools) | +16.60 |
跨应用与工具编排
Opus 4.7 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| PinchBench v2 | 76.0115 / 45Reported best (effort unspecified) | 75.7017 / 45Reported best (effort unspecified) | +0.31 |
跨能力综合测试
GPT-5.4 领先 1/1| 评测项 | Opus 4.7 | GPT-5.4 | 分差 |
|---|---|---|---|
| LiveBench | 76.5331 / 157Deep Thinking (No Tools) | 77.9722 / 157Deep Thinking (No Tools) | -1.44 |
测试模式不同的成绩
共 17 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 | Opus 4.7 | GPT-5.4 |
|---|---|---|
| Terminal Bench 2.0 | 69.40Extended (With Tools) | 75.10Extra-High (With Tools) |
| MedCode | 54.86Max (With Tools) | 41.29Extra-High (With Tools) |
| MedScribe | 82.95Max (With Tools) | 77.55Extra-High (With Tools) |
| Vibe Code Bench v1.1 | 71Max (With Tools) | 48.47Extra-High (With Tools) |
| OSWorld-Verified | 78Extended (With Tools) | 75Extra-High (With Tools) |
| BrowseComp | 79.30Extended (With Tools) | 82.70Extra-High (With Tools) |
| HLE | 54.70Extended (With Tools) | 52.10Extra-High (With Tools) |
| Code Migration | 43.88Max (With Tools) | 34.98Extra-High (With Tools) |
| FrontierMath Tier 4 v2 | 31.71Max (With Tools) | 49Extra-High (With Tools) |
| FrontierMath v2 | 70.18Max (With Tools) | 78.60Extra-High (With Tools) |
| EBR-bench | 19.05Max (With Tools) | 25.40Extra-High (With Tools) |
| Text Arena (Coding) | 1,562Normal (No Tools) | 1,457Thinking High (No Tools) |
| SWE-Bench Pro - Public | 64.30Extended (With Tools) | 57.70Extra-High (No Tools) |
| GPQA Diamond | 94.20Extended (No Tools) | 89.90Thinking High (No Tools) |
| SAGE | 56.10Max (With Tools) | 43.31Extra-High (With Tools) |
| τ³-Banking | 40.21Max (With Tools) | 39.43Extra-High (With Tools) |
| MCP-Atlas | 79.10Deep Thinking (With Tools) | 70.60Extra-High (With Tools) |
规格对比
| 字段 | Opus 4.7 | GPT-5.4 |
|---|---|---|
| 发布机构 | Anthropic | OpenAI |
| 发布时间 | 2026-04-16 | 2026-03-05 |
| 模型类型 | 推理大模型 | 多模态大模型 |
| 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 | 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 | Opus 4.7 | GPT-5.4 |
|---|---|---|
| 文本输入 | $5 / 1M tokens | $2.5 / 1M tokens |
| 文本输出 | $25 / 1M tokens | $15 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | $0.25 / 1M tokens |
| 缓存写入 | $6.25 / 1M tokens | 暂无公开价格 |
小结
- Opus 4.7在以下类目领先:仓库修复与多文件工程 (2/2)、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)、科学知识与推理 (1/1)、维护、优化与工程管理 (1/1)、自主开发与终端任务 (1/1)、跨应用与工具编排 (1/1)
- GPT-5.4在以下类目领先:数学 (5/5)、多轮记忆与持续上下文 (1/1)、跨能力综合测试 (1/1)
- 胶着类目:抽象归纳与泛化
18 个同模式 benchmark 上，Opus 4.7 平均高出 6.58 分。
单项差距最大的 benchmark：Creative Writing — Opus 4.7 1,914，GPT-5.4 1,840（分差 +74.60）。
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。