GPT-5.6 SolvsGPT-5.4
在 11 个同模式 benchmark 中，GPT-5.6 Sol 整体领先：GPT-5.6 Sol 领先 11 项，GPT-5.4 领先 0 项，持平 0 项，平均分差 +18.07。另有 16 项测试模式不同，仅供参考。
GPT-5.6 Sol
OpenAI · 2026-06-26 · 推理大模型
GPT-5.4
OpenAI · 2026-03-05 · 多模态大模型
GPT-5.6 Sol11 项(100%)(0%)0 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 11 项。
仓库修复与多文件工程
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| DeepSWE | 70.7321 / 100Extra-High (With Tools) | 51.7772 / 100Extra-High (With Tools) | +18.96 |
| APEX-SWE | 45.8028 / 73Extra-High (With Tools) | 34.3054 / 73Extra-High (With Tools) | +11.50 |
抽象归纳与泛化
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 97.5011 / 226Extra-High (No Tools) | 93.6746 / 226Extra-High (No Tools) | +3.83 |
| ARC-AGI-3 | 2.1519 / 60Thinking High (No Tools) | 0.2145 / 60Thinking High (No Tools) | +1.94 |
科学知识与推理
GPT-5.6 Sol 领先 2/2| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| GPQA Diamond | 89.9043 / 254Thinking Low (No Tools) | 84.8589 / 254Thinking Low (No Tools) | +5.05 |
| CritPt | 5.1098 / 204Normal (No Tools) | 0.60172 / 204Normal (No Tools) | +4.50 |
写作与创意表达
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| Creative Writing | 1,9729 / 112Normal (No Tools) | 1,84019 / 112Normal (No Tools) | +131.90 |
图像感知与视觉推理
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| MMMU-Pro | 71.90119 / 231Normal (No Tools) | 70.60126 / 231Normal (No Tools) | +1.30 |
客服与业务流程
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| τ³-Banking | 46.9113 / 168Extra-High (With Tools) | 39.4336 / 168Extra-High (With Tools) | +7.48 |
自主开发与终端任务
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| Terminal Bench Hard | 61.406 / 244Extra-High (With Tools) | 57.6011 / 244Extra-High (With Tools) | +3.80 |
跨应用与工具编排
GPT-5.6 Sol 领先 1/1| 评测项 | GPT-5.6 Sol | GPT-5.4 | 分差 |
|---|---|---|---|
| PinchBench v2 | 84.197 / 45Reported best (effort unspecified) | 75.7017 / 45Reported best (effort unspecified) | +8.49 |
测试模式不同的成绩
共 16 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-5.6 Sol | GPT-5.4 |
|---|---|---|
| MedCode | 43.97Max (With Tools) | 41.29Extra-High (With Tools) |
| MedScribe | 85.23Max (With Tools) | 77.55Extra-High (With Tools) |
| Vibe Code Bench v1.1 | 80.50Max (With Tools) | 48.47Extra-High (With Tools) |
| LiveBench | 81.05Max (No Tools) | 77.97Deep Thinking (No Tools) |
| BrowseComp | 90.40Reported best (effort unspecified) | 82.70Extra-High (With Tools) |
| HLE | 44.50Max (No Tools) | 52.10Extra-High (With Tools) |
| Code Migration | 52.92Max (With Tools) | 34.98Extra-High (With Tools) |
| FrontierMath Tier 4 v2 | 83Reported best (effort unspecified) | 49Extra-High (With Tools) |
| FrontierMath v2 | 89.12Max (With Tools) | 78.60Extra-High (With Tools) |
| Context Arena | 97.63Max (No Tools) | 32.79Normal (No Tools) |
| EBR-bench | 44.76Max (With Tools) | 25.40Extra-High (With Tools) |
| WeirdML v2 | 88.76Thinking High (With Tools) | 57.44Normal (With Tools) |
| Text Arena (Coding) | 1,620Extra-High (No Tools) | 1,457Thinking High (No Tools) |
| SWE-Bench Pro - Public | 64.60Extra-High (With Tools) | 57.70Extra-High (No Tools) |
| SAGE | 52.56Max (With Tools) | 43.31Extra-High (With Tools) |
| MCP-Atlas | 81.80Max (With Tools) | 70.60Extra-High (With Tools) |
规格对比
| 字段 | GPT-5.6 Sol | GPT-5.4 |
|---|---|---|
| 发布机构 | OpenAI | OpenAI |
| 发布时间 | 2026-06-26 | 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 | 1.05M | 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 | GPT-5.6 Sol | GPT-5.4 |
|---|---|---|
| 文本输入 | $4 / 1M tokens | $2.5 / 1M tokens |
| 文本输出 | $20 / 1M tokens | $15 / 1M tokens |
| 缓存读取 | $0.4 / 1M tokens | $0.25 / 1M tokens |
| 缓存写入 | $5 / 1M tokens | 暂无公开价格 |
小结
- GPT-5.6 Sol在以下类目领先:仓库修复与多文件工程 (2/2)、抽象归纳与泛化 (2/2)、科学知识与推理 (2/2)、写作与创意表达 (1/1)、图像感知与视觉推理 (1/1)、客服与业务流程 (1/1)、自主开发与终端任务 (1/1)、跨应用与工具编排 (1/1)
11 个同模式 benchmark 上，GPT-5.6 Sol 平均高出 18.07 分。
单项差距最大的 benchmark：Creative Writing — GPT-5.6 Sol 1,972，GPT-5.4 1,840（分差 +131.90）。
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。