GPT-5.4vsClaude Opus 4.6
在 13 个同模式 benchmark 中，Claude Opus 4.6 整体领先：GPT-5.4 领先 4 项，Claude Opus 4.6 领先 9 项，持平 0 项，平均分差 -3.01。另有 23 项测试模式不同，仅供参考。
GPT-5.4
OpenAI · 2026-03-05 · 多模态大模型
Claude Opus 4.6
Anthropic · 2026-02-05 · 推理大模型
GPT-5.44 项(31%)(69%)9 项Claude Opus 4.6
评测分数
按能力类目分组，每组内按分差大小排列；共 13 项。
仓库修复与多文件工程
Claude Opus 4.6 领先 2/2| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| APEX-SWE | 36.5047 / 73Thinking High (With Tools) | 40.5037 / 73Thinking High (With Tools) | -4 |
| FrontierSWE v1 | 45.9610 / 17Historical report (mode unspecified) | 48.909 / 17Historical report (mode unspecified) | -2.94 |
跨应用与工具编排
GPT-5.4 领先 2/2| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| PinchBench v2 | 75.7017 / 45Reported best (effort unspecified) | 69.9026 / 45Reported best (effort unspecified) | +5.80 |
| Pinch Bench | 90.501 / 38Thinking (With Tools) | 87.408 / 38Thinking (With Tools) | +3.10 |
写作与创意表达
GPT-5.4 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Creative Writing | 1,84019 / 112Normal (No Tools) | 1,80922 / 112Normal (No Tools) | +30.60 |
图像感知与视觉推理
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| MMMU-Pro | 70.60126 / 231Normal (No Tools) | 72.50117 / 231Normal (No Tools) | -1.90 |
多轮记忆与持续上下文
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Context Arena | 32.79112 / 126Normal (No Tools) | 60.3376 / 126Normal (No Tools) | -27.54 |
抽象归纳与泛化
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 86.17102 / 226Thinking Medium (No Tools) | 9258 / 226Thinking Medium (No Tools) | -5.83 |
机器学习工程
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| WeirdML v2 | 57.4432 / 53Normal (With Tools) | 65.9024 / 53Normal (With Tools) | -8.46 |
科学知识与推理
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| CritPt | 0.60172 / 204Normal (No Tools) | 2.80124 / 204Normal (No Tools) | -2.20 |
维护、优化与工程管理
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| GSO | 25.499 / 22Thinking High (With Tools) | 41.203 / 22Thinking High (With Tools) | -15.71 |
自主开发与终端任务
Claude Opus 4.6 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| Terminal Bench Hard | 37.9059 / 244Normal (With Tools) | 48.5025 / 244Normal (With Tools) | -10.60 |
跨能力综合测试
GPT-5.4 领先 1/1| 评测项 | GPT-5.4 | Claude Opus 4.6 | 分差 |
|---|---|---|---|
| LiveBench | 75.0744 / 157Thinking High (No Tools) | 74.5248 / 157Thinking High (No Tools) | +0.55 |
测试模式不同的成绩
共 23 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-5.4 | Claude Opus 4.6 |
|---|---|---|
| ARC-AGI-3 | 0.21Thinking High (No Tools) | 0.51Max (No Tools) |
| Terminal Bench 2.0 | 75.10Extra-High (With Tools) | 65.40Extended (With Tools) |
| MedCode | 41.29Extra-High (With Tools) | 48.24Normal (With Tools) |
| MedScribe | 77.55Extra-High (With Tools) | 86.74Normal (With Tools) |
| Vibe Code Bench v1.1 | 48.47Extra-High (With Tools) | 57.57Normal (With Tools) |
| OSWorld-Verified | 75Extra-High (With Tools) | 72.70Extended (With Tools) |
| BrowseComp | 82.70Extra-High (With Tools) | 84Thinking (With Tools + Internet) |
| HLE | 52.10Extra-High (With Tools) | 19Normal (No Tools) |
| AIME 2026 | 99.17Extra-High (With Tools) | 96.67Thinking High (With Tools) |
| FrontierMath | 47.60Extra-High (No Tools) | 40.70Max (No Tools) |
| FrontierMath - Tier 4 | 27.10Extra-High (No Tools) | 22.90Max (No Tools) |
| FrontierMath Tier 4 v2 | 49Extra-High (With Tools) | 26.83Max (With Tools) |
| FrontierMath v2 | 78.60Extra-High (With Tools) | 65.96Max (With Tools) |
| HMMT Feb 2026 | 97.73Extra-High (With Tools) | 96.21Thinking High (With Tools) |
| MathArena Apex | 54.17Extra-High (With Tools) | 34.45Thinking High (With Tools) |
| EBR-bench | 25.40Extra-High (With Tools) | 12.70Max (With Tools) |
| Text Arena (Coding) | 1,457Thinking High (No Tools) | 1,555Normal (No Tools) |
| SWE-Bench Pro - Commercial | 43.40Extra-High (With Tools) | 47.10Thinking (With Tools) |
| GPQA Diamond | 89.90Thinking High (No Tools) | 91.31Extended (No Tools) |
| SAGE | 43.31Extra-High (With Tools) | 51.58Max (With Tools) |
| τ²-Bench - Telecom | 64.30Normal (With Tools) | 99.25Extended (With Tools) |
| τ³-Banking | 39.43Extra-High (With Tools) | 27.32Max (With Tools) |
| MCP-Atlas | 70.60Extra-High (With Tools) | 76.80Deep Thinking (With Tools) |
规格对比
| 字段 | GPT-5.4 | Claude Opus 4.6 |
|---|---|---|
| 发布机构 | OpenAI | Anthropic |
| 发布时间 | 2026-03-05 | 2026-02-05 |
| 模型类型 | 多模态大模型 | 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 | 1000K |
| 最大输出 | 125K | 64K |
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-5.4 | Claude Opus 4.6 |
|---|---|---|
| 文本输入 | $2.5 / 1M tokens | $5 / 1M tokens |
| 文本输出 | $15 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | $0.25 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $6.25 / 1M tokens |
小结
- GPT-5.4在以下类目领先:跨应用与工具编排 (2/2)、写作与创意表达 (1/1)、跨能力综合测试 (1/1)
- Claude Opus 4.6在以下类目领先:仓库修复与多文件工程 (2/2)、图像感知与视觉推理 (1/1)、多轮记忆与持续上下文 (1/1)、抽象归纳与泛化 (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)
13 个同模式 benchmark 上，Claude Opus 4.6 平均高出 3.01 分。
单项差距最大的 benchmark：Creative Writing — GPT-5.4 1,840，Claude Opus 4.6 1,809（分差 +30.60）。
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。