Gemini 3.6 FlashvsGemini 3.5 Flash
在 30 个同模式 benchmark 中，Gemini 3.6 Flash 整体领先：Gemini 3.6 Flash 领先 16 项，Gemini 3.5 Flash 领先 14 项，持平 0 项，平均分差 +2.82。另有 2 项测试模式不同，仅供参考。
Gemini 3.6 Flash
Google DeepMind · 2026-07-21 · 多模态大模型
Gemini 3.5 Flash
Google DeepMind · 2026-05-19 · 多模态大模型
Gemini 3.6 Flash16 项(53%)(47%)14 项Gemini 3.5 Flash
评测分数
按能力类目分组，每组内按分差大小排列；共 30 项。
客服与业务流程
Gemini 3.5 Flash 领先 3/3| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Public Benefits Bench v1.1 | 56.5623 / 26Thinking High (With Tools) | 59.4720 / 26Thinking High (With Tools) | -2.91 |
| τ³-Banking | 29.9062 / 171Thinking High (With Tools) | 32.2055 / 171Thinking High (With Tools) | -2.30 |
| SAGE | 48.6918 / 65Thinking High (With Tools) | 49.8811 / 65Thinking High (With Tools) | -1.19 |
法律
Gemini 3.6 Flash 领先 2/3| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Legal Research Bench | 2538 / 47Thinking High (With Tools) | 30.7735 / 47Thinking High (With Tools) | -5.77 |
| Harvey Lab-AA | 85.1521 / 44Thinking High (With Tools) | 82.0828 / 44Thinking High (With Tools) | +3.07 |
| Harvey's Legal Agent Benchmark | 3.3333 / 51Thinking High (With Tools) | 2.5036 / 51Thinking High (With Tools) | +0.83 |
临床工作与医疗决策
胶着 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| MedScribe | 79.6643 / 68Thinking High (With Tools) | 76.5751 / 68Thinking High (With Tools) | +3.09 |
| MedCode | 53.159 / 66Thinking High (With Tools) | 55.834 / 66Thinking High (With Tools) | -2.68 |
仓库修复与多文件工程
Gemini 3.6 Flash 领先 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| DeepSWE | 46.6882 / 106Thinking High (With Tools) | 36.0694 / 106Thinking High (With Tools) | +10.62 |
| APEX-SWE | 39.4040 / 73Thinking High (With Tools) | 36.1050 / 73Thinking High (With Tools) | +3.30 |
数学
Gemini 3.5 Flash 领先 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| FrontierMath Tier 4 v2 | 21.9547 / 71Thinking High (With Tools) | 26.8340 / 71Thinking High (With Tools) | -4.88 |
| FrontierMath v2 | 58.9537 / 114Thinking High (With Tools) | 62.8134 / 114Thinking High (With Tools) | -3.86 |
文档与图表理解
胶着 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| CharXiv RQ | 89.404 / 25Thinking (With Tools) | 84.9011 / 25Thinking (With Tools) | +4.50 |
| GDP.pdf | 17.4053 / 126Thinking High (No Tools) | 19.8045 / 126Thinking High (No Tools) | -2.40 |
科学知识与推理
胶着 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| CritPt | 10.6077 / 208Thinking High (No Tools) | 13.1067 / 208Thinking High (No Tools) | -2.50 |
| GPQA Diamond | 94.138 / 258Thinking High (No Tools) | 92.8020 / 258Thinking High (No Tools) | +1.33 |
金融
胶着 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| EMB (Excel Modeling) | 65.4121 / 47Thinking High (With Tools) | 63.5524 / 47Thinking High (With Tools) | +1.86 |
| Finance Agent v2 | 56.3017 / 47Thinking High (With Tools) | 57.8611 / 47Thinking High (With Tools) | -1.56 |
长文定位与检索
Gemini 3.6 Flash 领先 2/2| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| GDM-MRCR v2 (8-needle, 1M) | 541 / 3Thinking (No Tools) | 26.602 / 3Thinking (No Tools) | +27.40 |
| GDM-MRCR v2 (8-needle, 128K) | 91.802 / 8Thinking (No Tools) | 77.304 / 8Thinking (No Tools) | +14.50 |
图像感知与视觉推理
Gemini 3.5 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| MMMU-Pro | 83.2018 / 233Thinking High (No Tools) | 84.3012 / 233Thinking High (No Tools) | -1.10 |
多轮记忆与持续上下文
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Context Arena | 88.7817 / 126Thinking High (No Tools) | 77.1947 / 126Thinking High (No Tools) | +11.59 |
抽象归纳与泛化
Gemini 3.5 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 91.1764 / 226Thinking High (With Tools) | 92.5053 / 226Thinking High (With Tools) | -1.33 |
机器学习工程
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| MLE-Bench | 63.901 / 3Thinking (With Tools) | 49.702 / 3Thinking (With Tools) | +14.20 |
科学计算与科研编程
Gemini 3.5 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| SciCode | 53.4055 / 137Thinking High (No Tools) | 53.9052 / 137Thinking High (No Tools) | -0.50 |
程序生成与编辑
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Vibe Code Bench v1.1 | 6435 / 67Thinking High (With Tools) | 48.6842 / 67Thinking High (With Tools) | +15.32 |
维护、优化与工程管理
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Code Migration | 30.9333 / 49Thinking High (With Tools) | 26.7537 / 49Thinking High (With Tools) | +4.18 |
自主开发与终端任务
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 7.1071 / 108Thinking High (With Tools) | 6.6074 / 108Thinking High (With Tools) | +0.50 |
跨能力综合测试
Gemini 3.5 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| LiveBench | 73.5952 / 157Thinking High (No Tools) | 74.6446 / 157Thinking High (No Tools) | -1.05 |
跨能力聚合指数
Gemini 3.6 Flash 领先 1/1| 评测项 | Gemini 3.6 Flash | Gemini 3.5 Flash | 分差 |
|---|---|---|---|
| Vals Index | 55.3525 / 42Thinking High (With Tools) | 53.0828 / 42Thinking High (With Tools) | +2.27 |
测试模式不同的成绩
共 2 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 3.6 Flash | Gemini 3.5 Flash |
|---|---|---|
| OSWorld-Verified | 83Thinking (With Tools) | 78.40Thinking High (With Tools) |
| SWE-Bench Pro - Public | 58.70Thinking (With Tools) | 55.10Thinking High (With Tools) |
规格对比
| 字段 | Gemini 3.6 Flash | Gemini 3.5 Flash |
|---|---|---|
| 发布机构 | Google DeepMind | Google DeepMind |
| 发布时间 | 2026-07-21 | 2026-05-19 |
| 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 | 64K | 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 | Gemini 3.6 Flash | Gemini 3.5 Flash |
|---|---|---|
| 文本输入 | $1.5 / 1M tokens | $1.5 / 1M tokens |
| 文本输出 | $7.5 / 1M tokens | $9 / 1M tokens |
| 缓存读取 | $0.15 / 1M tokens | $0.15 / 1M tokens |
小结
- Gemini 3.6 Flash在以下类目领先:法律 (2/3)、仓库修复与多文件工程 (2/2)、长文定位与检索 (2/2)、多轮记忆与持续上下文 (1/1)、机器学习工程 (1/1)、程序生成与编辑 (1/1)、维护、优化与工程管理 (1/1)、自主开发与终端任务 (1/1)、跨能力聚合指数 (1/1)
- Gemini 3.5 Flash在以下类目领先:客服与业务流程 (3/3)、数学 (2/2)、图像感知与视觉推理 (1/1)、抽象归纳与泛化 (1/1)、科学计算与科研编程 (1/1)、跨能力综合测试 (1/1)
- 胶着类目:临床工作与医疗决策、文档与图表理解、科学知识与推理、金融
30 个同模式 benchmark 上，Gemini 3.6 Flash 平均高出 2.82 分。
单项差距最大的 benchmark：GDM-MRCR v2 (8-needle, 1M) — Gemini 3.6 Flash 54，Gemini 3.5 Flash 26.60（分差 +27.40）。
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。