Gemini 3.8 FlashvsGPT-5.6 Terra
在 10 个同模式 benchmark 中，Gemini 3.8 Flash 整体领先：Gemini 3.8 Flash 领先 8 项，GPT-5.6 Terra 领先 2 项，持平 0 项，平均分差 -2.75。另有 26 项测试模式不同，仅供参考。
Gemini 3.8 Flash
Google DeepMind · 2026-09-02 · 多模态大模型
GPT-5.6 Terra
OpenAI · 2026-06-26 · 推理大模型
Gemini 3.8 Flash8 项(80%)(20%)2 项GPT-5.6 Terra
评测分数
按能力类目分组，每组内按分差大小排列；共 10 项。
自主开发与终端任务
Gemini 3.8 Flash 领先 2/2| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 19.7045 / 105Thinking High (With Tools) | 1.5090 / 105Thinking High (With Tools) | +18.20 |
| CursorBench 4.0 | 39.6016 / 43Thinking High (With Tools) | 30.7032 / 43Thinking High (With Tools) | +8.90 |
仓库修复与多文件工程
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| DeepSWE | 73.839 / 101Thinking High (With Tools) | 53.7668 / 101Thinking High (With Tools) | +20.07 |
写作与创意表达
GPT-5.6 Terra 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| Creative Writing | 1,74826 / 112Normal (No Tools) | 1,85514 / 112Normal (No Tools) | -107 |
图像感知与视觉推理
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| MMMU-Pro | 85.604 / 231Thinking High (No Tools) | 79.1051 / 231Thinking High (No Tools) | +6.50 |
客服与业务流程
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| τ³-Banking | 44.9017 / 168Thinking High (With Tools) | 28.7067 / 168Thinking High (With Tools) | +16.20 |
抽象归纳与泛化
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| ARC-AGI-3 | 10.3712 / 60Thinking High (No Tools) | 0.4932 / 60Thinking High (No Tools) | +9.88 |
文档与图表理解
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| GDP.pdf | 2135 / 124Thinking High (No Tools) | 20.8038 / 124Thinking High (No Tools) | +0.20 |
科学知识与推理
GPT-5.6 Terra 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| CritPt | 18.3045 / 204Thinking High (No Tools) | 22.9033 / 204Thinking High (No Tools) | -4.60 |
科学计算与科研编程
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | GPT-5.6 Terra | 分差 |
|---|---|---|---|
| SciCode | 56.6025 / 135Thinking High (No Tools) | 52.4059 / 135Thinking High (No Tools) | +4.20 |
测试模式不同的成绩
共 26 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 | Gemini 3.8 Flash | GPT-5.6 Terra |
|---|---|---|
| ARC-AGI-1 | 98.50Thinking High (No Tools) | 96.50Max (No Tools) |
| Terminal-Bench 2.1 | 89.40Thinking (With Tools) | 87.40Max (No Tools) |
| IOI (Vals v2) | 56.94Thinking High (With Tools) | 87.61Max (With Tools) |
| Vals Index | 62.25Thinking High (With Tools) | 65.14Extra-High (No Tools) |
| Vals Index v2.1 | 54.83Thinking High (With Tools) | 53.09Max (task-specific tools) |
| MedCode | 48.13Thinking High (With Tools) | 43.41Max (With Tools) |
| MedScribe | 84.50Thinking High (With Tools) | 82.87Max (With Tools) |
| Vibe Code Bench v1.1 | 78.65Thinking High (With Tools) | 74.59Max (With Tools) |
| SimpleBench | 82.40Thinking Medium (No Tools) | 48.90Extra-High (No Tools) |
| LiveBench | 75.83Thinking High (No Tools) | 77.94Max (No Tools) |
| EMB (Excel Modeling) | 72.20Thinking High (With Tools) | 66.20Max (With Tools) |
| Finance Agent v2 | 61.40Thinking (With Tools) | 54.44Max (With Tools) |
| Tax Agent Bench | 66.77Thinking High (With Tools) | 65.20Max (With Tools) |
| HLE-Verified | 54.90Thinking Medium (No Tools) | 51.10Reported best (effort unspecified) |
| Harvey's Legal Agent Benchmark | 10Thinking High (With Tools) | 0.83Max (With Tools) |
| Legal Research Bench | 38.94Thinking High (With Tools) | 41.35Max (With Tools) |
| Code Migration | 36.55Thinking High (With Tools) | 47.80Max (With Tools) |
| FrontierMath Tier 4 v2 | 21.95Thinking High (With Tools) | 70.73Max (With Tools) |
| FrontierMath v2 | 68.42Thinking High (With Tools) | 85.96Max (With Tools) |
| ProofBench v1.1 (Lean 4) | 48Thinking High (With Tools) | 74Max (With Tools) |
| APEX-SWE | 36.30Thinking High (With Tools) | 31.10Max (With Tools) |
| SWE-Bench Pro V2 | 94.86Thinking High (With Tools) | 92.37Extra-High (With Tools) |
| SWE-Bench Pro V2 Hard | 58.80Thinking High (With Tools) | 86.30Extra-High (With Tools) |
| Terminal-Bench-Science 0.1 | 12.40High Thinking (With Tools) | 8.60Max (With Tools) |
| Public Benefits Bench v1.1 | 65.29Thinking High (With Tools) | 62.38Max (With Tools) |
| SAGE | 35.06Thinking High (With Tools) | 47Max (With Tools) |
规格对比
| 字段 | Gemini 3.8 Flash | GPT-5.6 Terra |
|---|---|---|
| 发布机构 | Google DeepMind | OpenAI |
| 发布时间 | 2026-09-02 | 2026-06-26 |
| 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 | 1.05M |
| 最大输出 | 64K | 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 | Gemini 3.8 Flash | GPT-5.6 Terra |
|---|---|---|
| 文本输入 | $0.75 / 1M tokens | $2 / 1M tokens |
| 文本输出 | $3.75 / 1M tokens | $12 / 1M tokens |
| 缓存读取 | $0.075 / 1M tokens | $0.2 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $2.5 / 1M tokens |
小结
- Gemini 3.8 Flash在以下类目领先:自主开发与终端任务 (2/2)、仓库修复与多文件工程 (1/1)、图像感知与视觉推理 (1/1)、客服与业务流程 (1/1)、抽象归纳与泛化 (1/1)、文档与图表理解 (1/1)、科学计算与科研编程 (1/1)
- GPT-5.6 Terra在以下类目领先:写作与创意表达 (1/1)、科学知识与推理 (1/1)
10 个同模式 benchmark 上，GPT-5.6 Terra 平均高出 2.75 分。
单项差距最大的 benchmark：Creative Writing — Gemini 3.8 Flash 1,748，GPT-5.6 Terra 1,855（分差 -107）。
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。