GPT-5.1vsGPT-5
在 23 个共同 benchmark 中，GPT-5 整体领先：GPT-5.1 领先 8 项，GPT-5 领先 13 项，持平 2 项，平均分差 -6.28。
GPT-5.18 项(35%)持平2(57%)13 项GPT-5
评测分数
按能力类目分组，每组内按分差大小排列；共 23 项。
Coding and Software Engineer
GPT-5.1 领先 4/6| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| SWE-Bench Pro - Public | 50.8046 / 62Thinking High (No Tools) | 36.3059 / 62Thinking High (No Tools) | +14.50 |
| Text Arena (Coding) | 1,38732 / 35Thinking Medium (No Tools) | 1,39529 / 35Thinking Medium (No Tools) | -8 |
| GSO | 13.7013 / 21Thinking High (With Tools) | 6.9016 / 21Thinking High (With Tools) | +6.80 |
| LiveCodeBench | 49.40167 / 250Normal (No Tools) | 55.80149 / 250Normal (No Tools) | -6.40 |
| SWE-bench Verified | 76.3034 / 116Thinking High (No Tools) | 72.8051 / 116Thinking High (No Tools) | +3.50 |
| WeirdML v2 | 60.7728 / 52Thinking High (With Tools) | 60.7029 / 52Thinking High (With Tools) | +0.07 |
Math and Reasoning
GPT-5 领先 2/5| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| AIME2025 | 38170 / 215Normal (No Tools) | 61.90135 / 215Normal (No Tools) | -23.90 |
| IMO-ProofBench Advanced | 7.1016 / 24Thinking (No Tools) | 2011 / 24Thinking (No Tools) | -12.90 |
| FrontierMath | 26.7013 / 60Thinking High (With Tools) | 26.3014 / 60Thinking High (With Tools) | +0.40 |
| FrontierMath - Tier 4 | 12.5029 / 80Thinking High (No Tools) | 12.5029 / 80Thinking High (No Tools) | 持平 |
| MathArena Apex | 1.0414 / 17Thinking High (With Tools) | 1.0414 / 17Thinking High (With Tools) | 持平 |
Agent Level Benchmark
GPT-5 领先 2/3| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| τ²-Bench - Telecom | 46.50178 / 264Normal (With Tools) | 67144 / 264Normal (With Tools) | -20.50 |
| τ³-Banking | 15.90103 / 164Thinking High (With Tools) | 22.1085 / 164Thinking High (With Tools) | -6.20 |
| Terminal Bench Hard | 22.70143 / 244Normal (With Tools) | 18.20154 / 244Normal (With Tools) | +4.50 |
General Knowledge
GPT-5 领先 2/2| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| HLE | 5.30471 / 563Normal (No Tools) · Text only | 6.60450 / 563Normal (No Tools) · Text only | -1.30 |
| CritPt | 4.9098 / 200Thinking High (No Tools) | 5.7088 / 200Thinking High (No Tools) | -0.80 |
Multimodal Understanding
胶着 2/2| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| VPCT | 58.707 / 24Thinking High (No Tools) | 665 / 24Thinking High (No Tools) | -7.30 |
| MMMU-Pro | 62.40163 / 227Normal (No Tools) | 62.10165 / 227Normal (No Tools) | +0.30 |
AI Agent - Tool Usage
GPT-5.1 领先 1/1| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| Terminal-Bench 2.1 | 52.40126 / 192Thinking High (With Tools) | 35.20151 / 192Thinking High (With Tools) | +17.20 |
Instruction Following
GPT-5 领先 1/1| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| IF Bench | 43.20192 / 282Normal (No Tools) | 45.60178 / 282Normal (No Tools) | -2.40 |
Productivity Knowledge
GPT-5 领先 1/1| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| GDPval-AA v2 | 93093 / 105Thinking High (With Tools) | 1,01589 / 105Thinking High (With Tools) | -85 |
常识推理
GPT-5 领先 1/1| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| SimpleBench | 53.2045 / 93Thinking High (No Tools) | 56.7040 / 93Thinking High (No Tools) | -3.50 |
科学与综合推理
GPT-5 领先 1/1| 评测项 | GPT-5.1 | GPT-5 | 分差 |
|---|---|---|---|
| GPQA Diamond | 64.30361 / 462Normal (No Tools) | 77.80257 / 462Normal (No Tools) | -13.50 |
规格对比
| 字段 | GPT-5.1 | GPT-5 |
|---|---|---|
| 发布机构 | OpenAI | OpenAI |
| 发布时间 | 2025-11-12 | 2025-08-07 |
| 模型类型 | 推理大模型 | 基础大模型 |
| 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 | 400K | 400K |
| 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-5.1 | GPT-5 |
|---|---|---|
| 文本输入 | $1.25 / 1M tokens | $1.25 / 1M tokens |
| 文本输出 | $10 / 1M tokens | $10 / 1M tokens |
| 缓存读取 | $0.125 / 1M tokens | $0.125 / 1M tokens |
| 缓存写入 | $0 / 1M tokens | $0 / 1M tokens |
小结
- GPT-5.1在以下类目领先:Coding and Software Engineer (4/6)、AI Agent - Tool Usage (1/1)
- GPT-5在以下类目领先:Math and Reasoning (2/5)、Agent Level Benchmark (2/3)、General Knowledge (2/2)、Instruction Following (1/1)、Productivity Knowledge (1/1)、常识推理 (1/1)、科学与综合推理 (1/1)
- 胶着类目:Multimodal Understanding
23 个共同 benchmark 上，GPT-5 平均高出 6.28 分。
单项差距最大的 benchmark：GDPval-AA v2 — GPT-5.1 930，GPT-5 1,015（分差 -85）。
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。