GPT-5.4vsGemini 3.1 Pro Preview
在 10 个同模式 benchmark 中，Gemini 3.1 Pro Preview 整体领先：GPT-5.4 领先 4 项，Gemini 3.1 Pro Preview 领先 6 项，持平 0 项，平均分差 +33.72。另有 24 项测试模式不同，仅供参考。
GPT-5.4
OpenAI · 2026-03-05 · 多模态大模型
Gemini 3.1 Pro Preview
Google DeepMind · 2026-02-20 · 多模态大模型
GPT-5.44 项(40%)(60%)6 项Gemini 3.1 Pro Preview
评测分数
按能力类目分组，每组内按分差大小排列；共 10 项。
仓库修复与多文件工程
GPT-5.4 领先 2/2| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| FrontierSWE v1 | 45.9610 / 17Historical report (mode unspecified) | 34.3811 / 17Historical report (mode unspecified) | +11.58 |
| APEX-SWE | 36.5047 / 73Thinking High (With Tools) | 33.9055 / 73Thinking High (With Tools) | +2.60 |
跨应用与工具编排
胶着 2/2| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| PinchBench v2 | 75.7017 / 45Reported best (effort unspecified) | 81.019 / 45Reported best (effort unspecified) | -5.31 |
| Pinch Bench | 90.501 / 38Thinking (With Tools) | 86.7011 / 38Thinking (With Tools) | +3.80 |
写作与创意表达
GPT-5.4 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| Creative Writing | 1,84019 / 112Normal (No Tools) | 1,49157 / 112Normal (No Tools) | +348.40 |
图像感知与视觉推理
Gemini 3.1 Pro Preview 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| MMMU-Pro | 81.2032 / 231Thinking (No Tools) | 82.4022 / 231Thinking (No Tools) | -1.20 |
多轮记忆与持续上下文
Gemini 3.1 Pro Preview 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| Context Arena | 82.9928 / 126Thinking High (No Tools) | 84.7325 / 126Thinking High (No Tools) | -1.74 |
机器学习工程
Gemini 3.1 Pro Preview 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| WeirdML v2 | 57.4432 / 53Normal (With Tools) | 72.1017 / 53Normal (With Tools) | -14.66 |
科学知识与推理
Gemini 3.1 Pro Preview 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| GPQA Diamond | 89.9043 / 253Thinking High (No Tools) | 94.306 / 253Thinking High (No Tools) | -4.40 |
跨能力综合测试
Gemini 3.1 Pro Preview 领先 1/1| 评测项 | GPT-5.4 | Gemini 3.1 Pro Preview | 分差 |
|---|---|---|---|
| LiveBench | 75.0744 / 157Thinking High (No Tools) | 76.9529 / 157Thinking High (No Tools) | -1.88 |
测试模式不同的成绩
共 24 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 | Gemini 3.1 Pro Preview |
|---|---|---|
| Terminal Bench 2.0 | 75.10Extra-High (With Tools) | 68.50Thinking High (With Tools) |
| Terminal Bench Hard | 37.90Normal (With Tools) | 53.80Thinking (With Tools) |
| MedCode | 41.29Extra-High (With Tools) | 59.06Thinking High (With Tools) |
| MedScribe | 77.55Extra-High (With Tools) | 76.11Thinking High (With Tools) |
| Vibe Code Bench v1.1 | 48.47Extra-High (With Tools) | 32.03Thinking High (With Tools) |
| OSWorld-Verified | 75Extra-High (With Tools) | 76.20Thinking (With Tools) |
| BrowseComp | 82.70Extra-High (With Tools) | 85.90Thinking High (With Tools + Internet) |
| HLE | 52.10Extra-High (With Tools) | 51.40Thinking High (With Tools) |
| Code Migration | 34.98Extra-High (With Tools) | 17.31Thinking High (With Tools) |
| GSO | 31.37Extra-High (With Tools) | 22.55Normal (With Tools) |
| FrontierMath | 47.60Extra-High (No Tools) | 36.90Thinking High (No Tools) |
| FrontierMath - Tier 4 | 27.10Extra-High (No Tools) | 16.70Normal (No Tools) |
| FrontierMath Tier 4 v2 | 49Extra-High (With Tools) | 26.83Reported best (effort unspecified) |
| FrontierMath v2 | 78.60Extra-High (With Tools) | 59.65Reported best (effort unspecified) |
| EBR-bench | 25.40Extra-High (With Tools) | 14.29Thinking (With Tools) |
| Text Arena (Coding) | 1,457Thinking High (No Tools) | 1,461Normal (No Tools) |
| DeepSWE | 51.77Extra-High (With Tools) | 11.73Thinking High (With Tools) |
| SWE-Bench Pro - Commercial | 43.40Extra-High (With Tools) | 32.20Thinking (With Tools) |
| SWE-Bench Pro - Public | 57.70Extra-High (No Tools) | 54.20Thinking High (With Tools) |
| CritPt | 0.60Normal (No Tools) | 17.70Thinking (No Tools) |
| SAGE | 43.31Extra-High (With Tools) | 48.68Thinking High (With Tools) |
| τ²-Bench - Telecom | 64.30Normal (With Tools) | 99.30Thinking High (With Tools) |
| τ³-Banking | 39.43Extra-High (With Tools) | 26.03Thinking High (With Tools) |
| MCP-Atlas | 70.60Extra-High (With Tools) | 78.20Thinking High (With Tools) |
规格对比
| 字段 | GPT-5.4 | Gemini 3.1 Pro Preview |
|---|---|---|
| 发布机构 | OpenAI | Google DeepMind |
| 发布时间 | 2026-03-05 | 2026-02-20 |
| 模型类型 | 多模态大模型 | 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 | 1M | 1M |
| 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 | Gemini 3.1 Pro Preview |
|---|---|---|
| 文本输入 | $2.5 / 1M tokens | $2 / 1M tokens |
| 文本输出 | $15 / 1M tokens | $12 / 1M tokens |
| 缓存读取 | $0.25 / 1M tokens | $0.2 / 1M tokens |
小结
- GPT-5.4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)
- Gemini 3.1 Pro Preview在以下类目领先:图像感知与视觉推理 (1/1)、多轮记忆与持续上下文 (1/1)、机器学习工程 (1/1)、科学知识与推理 (1/1)、跨能力综合测试 (1/1)
- 胶着类目:跨应用与工具编排
10 个同模式 benchmark 上，GPT-5.4 平均高出 33.72 分。
单项差距最大的 benchmark：Creative Writing — GPT-5.4 1,840，Gemini 3.1 Pro Preview 1,491（分差 +348.40）。
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。