Gemini 3.8 FlashvsClaude Sonnet 5
在 8 个同模式 benchmark 中，Gemini 3.8 Flash 整体领先：Gemini 3.8 Flash 领先 6 项，Claude Sonnet 5 领先 2 项，持平 0 项，平均分差 +1.53。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
Claude Sonnet 5
Anthropic · 2026-06-30 · 多模态大模型
Gemini 3.8 Flash6 项(75%)(25%)2 项Claude Sonnet 5
评测分数
按能力类目分组，每组内按分差大小排列；共 8 项。
仓库修复与多文件工程
胶着 2/2| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| DeepSWE | 73.839 / 101Thinking High (With Tools) | 48.2378 / 101Thinking High (With Tools) | +25.60 |
| APEX-SWE | 36.3049 / 73Thinking High (With Tools) | 43.6532 / 73Thinking High (With Tools) | -7.35 |
自主开发与终端任务
Gemini 3.8 Flash 领先 2/2| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 19.7045 / 105Thinking High (With Tools) | 5.1074 / 105Thinking High (With Tools) | +14.60 |
| CursorBench 4.0 | 39.6016 / 43Thinking High (With Tools) | 30.8031 / 43Thinking High (With Tools) | +8.80 |
写作与创意表达
Claude Sonnet 5 领先 1/1| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| Creative Writing | 1,74826 / 112Normal (No Tools) | 1,79424 / 112Normal (No Tools) | -46.10 |
文档与图表理解
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| GDP.pdf | 22.8028 / 124Thinking Medium (No Tools) | 11.6078 / 124Thinking Medium (No Tools) | +11.20 |
科学知识与推理
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| CritPt | 18.3045 / 204Thinking High (No Tools) | 15.1058 / 204Thinking High (No Tools) | +3.20 |
科学计算与科研编程
Gemini 3.8 Flash 领先 1/1| 评测项 | Gemini 3.8 Flash | Claude Sonnet 5 | 分差 |
|---|---|---|---|
| SciCode | 56.6025 / 135Thinking High (No Tools) | 54.3046 / 135Thinking High (No Tools) | +2.30 |
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 | Claude Sonnet 5 |
|---|---|---|
| Terminal-Bench 2.1 | 89.40Thinking (With Tools) | 80.40Extra-High (With Tools) |
| IOI (Vals v2) | 56.94Thinking High (With Tools) | 45Max (With Tools) |
| Vals Index | 62.25Thinking High (With Tools) | 59.61Max (With Tools) |
| Vals Index v2.1 | 54.83Thinking High (With Tools) | 51.77Max (task-specific tools) |
| MedCode | 48.13Thinking High (With Tools) | 47.54Max (With Tools) |
| MedScribe | 84.50Thinking High (With Tools) | 76.05Max (With Tools) |
| Vibe Code Bench v1.1 | 78.65Thinking High (With Tools) | 81.33Max (With Tools) |
| SimpleBench | 82.40Thinking Medium (No Tools) | 60.60Normal (No Tools) |
| LiveBench | 75.83Thinking High (No Tools) | 76.04Extra-High (No Tools) |
| EMB (Excel Modeling) | 72.20Thinking High (With Tools) | 66.32Max (With Tools) |
| Finance Agent v2 | 61.40Thinking (With Tools) | 53.91Max (With Tools) |
| Tax Agent Bench | 66.77Thinking High (With Tools) | 62.27Max (With Tools) |
| HLE | 44.52Reported best (effort unspecified) | 57.40Extra-High (With Tools) |
| HLE-Verified | 54.90Thinking Medium (No Tools) | 31Reported best (effort unspecified) |
| Harvey's Legal Agent Benchmark | 10Thinking High (With Tools) | 5Max (With Tools) |
| Legal Research Bench | 38.94Thinking High (With Tools) | 41.83Max (With Tools) |
| Code Migration | 36.55Thinking High (With Tools) | 44.39Max (With Tools) |
| FrontierMath Tier 4 v2 | 21.95Thinking High (With Tools) | 29.27Max (With Tools) |
| FrontierMath v2 | 68.42Thinking High (With Tools) | 65.61Max (With Tools) |
| ProofBench v1.1 (Lean 4) | 48Thinking High (With Tools) | 77Max (With Tools) |
| SWE-Bench Pro V2 | 94.86Thinking High (With Tools) | 93.15Extra-High (With Tools) |
| SWE-Bench Pro V2 Hard | 58.80Thinking High (With Tools) | 88.20Extra-High (With Tools) |
| Public Benefits Bench v1.1 | 65.29Thinking High (With Tools) | 66.03Max (With Tools) |
| SAGE | 35.06Thinking High (With Tools) | 48.92Max (With Tools) |
| τ³-Banking | 45.80Thinking Medium (With Tools) | 15.70Normal (With Tools) |
| MMMU-Pro | 85.60Thinking High (No Tools) | 71.90Normal (No Tools) |
规格对比
| 字段 | Gemini 3.8 Flash | Claude Sonnet 5 |
|---|---|---|
| 发布机构 | Google DeepMind | Anthropic |
| 发布时间 | 2026-09-02 | 2026-06-30 |
| 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 | 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 | Claude Sonnet 5 |
|---|---|---|
| 文本输入 | $0.75 / 1M tokens | $2 / 1M tokens |
| 文本输出 | $3.75 / 1M tokens | $10 / 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)、文档与图表理解 (1/1)、科学知识与推理 (1/1)、科学计算与科研编程 (1/1)
- Claude Sonnet 5在以下类目领先:写作与创意表达 (1/1)
- 胶着类目:仓库修复与多文件工程
8 个同模式 benchmark 上，Gemini 3.8 Flash 平均高出 1.53 分。
单项差距最大的 benchmark：Creative Writing — Gemini 3.8 Flash 1,748，Claude Sonnet 5 1,794（分差 -46.10）。
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。