Gemini 3.5 Flash-LitevsGPT-5.6 Luna
在 3 个同模式 benchmark 中，GPT-5.6 Luna 整体领先：Gemini 3.5 Flash-Lite 领先 0 项，GPT-5.6 Luna 领先 3 项，持平 0 项，平均分差 -99.72。另有 20 项测试模式不同，仅供参考。
Gemini 3.5 Flash-Lite
Google DeepMind · 2026-07-21 · 多模态大模型
GPT-5.6 Luna
OpenAI · 2026-06-26 · 推理大模型
Gemini 3.5 Flash-Lite0 项(0%)(100%)3 项GPT-5.6 Luna
评测分数
按能力类目分组，每组内按分差大小排列；共 3 项。
写作与创意表达
GPT-5.6 Luna 领先 1/1| 评测项 | Gemini 3.5 Flash-Lite | GPT-5.6 Luna | 分差 |
|---|---|---|---|
| Creative Writing | 1,55948 / 112Normal (No Tools) | 1,82920 / 112Normal (No Tools) | -269.60 |
抽象归纳与泛化
GPT-5.6 Luna 领先 1/1| 评测项 | Gemini 3.5 Flash-Lite | GPT-5.6 Luna | 分差 |
|---|---|---|---|
| ARC-AGI-1 | 53.50163 / 221Thinking High (With Tools) | 76.50119 / 221Thinking High (With Tools) | -23 |
科学知识与推理
GPT-5.6 Luna 领先 1/1| 评测项 | Gemini 3.5 Flash-Lite | GPT-5.6 Luna | 分差 |
|---|---|---|---|
| GPQA Diamond | 75.76152 / 253Thinking Low (No Tools) | 82.32114 / 253Thinking Low (No Tools) | -6.57 |
测试模式不同的成绩
共 20 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 3.5 Flash-Lite | GPT-5.6 Luna |
|---|---|---|
| Terminal-Bench 2.1 | 54Thinking (With Tools) | 84.70Max (No Tools) |
| Terminal-Bench 4.0 | 1Thinking (With Tools) | 17.27Max (With Tools) |
| Vals Index | 36.71Thinking High (With Tools) | 59.88Max (With Tools) |
| MedCode | 43.49Thinking High (With Tools) | 42.39Max (With Tools) |
| MedScribe | 70.89Thinking High (With Tools) | 84.39Max (With Tools) |
| Vibe Code Bench v1.1 | 37.16Thinking High (With Tools) | 77.06Max (With Tools) |
| LiveBench | 63.94Thinking High (No Tools) | 73.56Max (No Tools) |
| GDP.pdf | 13.60Thinking (No Tools) | 24Max (No Tools) |
| EMB (Excel Modeling) | 43.23Thinking High (With Tools) | 67.12Max (With Tools) |
| Finance Agent v2 | 47.44Thinking High (With Tools) | 55.04Max (With Tools) |
| Legal Research Bench | 13.94Thinking High (With Tools) | 36.54Max (With Tools) |
| Code Migration | 6.15Thinking High (With Tools) | 44.55Max (With Tools) |
| FrontierMath v2 | 25.96Thinking High (With Tools) | 39.65Normal (No Tools) |
| Context Arena | 72.57Thinking High (No Tools) | 81.80Max (No Tools) |
| APEX-SWE | 31.20Thinking High (With Tools) | 32.10Max (With Tools) |
| SciCode | 41.30Thinking (No Tools) | 53.60Max (No Tools) |
| Public Benefits Bench v1.1 | 51.83Thinking High (With Tools) | 61.16Max (With Tools) |
| SAGE | 47.30Thinking High (With Tools) | 44.22Max (With Tools) |
| τ³-Banking | 17.50Thinking (With Tools) | 9.70Normal (With Tools) |
| MMMU-Pro | 79Thinking (No Tools) | 60.30Normal (No Tools) |
规格对比
| 字段 | Gemini 3.5 Flash-Lite | GPT-5.6 Luna |
|---|---|---|
| 发布机构 | Google DeepMind | OpenAI |
| 发布时间 | 2026-07-21 | 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 3.5 Flash-Lite | GPT-5.6 Luna |
|---|---|---|
| 文本输入 | $0.3 / 1M tokens | $0.2 / 1M tokens |
| 文本输出 | $2.5 / 1M tokens | $1.2 / 1M tokens |
| 缓存读取 | $0.03 / 1M tokens | $0.02 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $0.25 / 1M tokens |
小结
- GPT-5.6 Luna在以下类目领先:写作与创意表达 (1/1)、抽象归纳与泛化 (1/1)、科学知识与推理 (1/1)
3 个同模式 benchmark 上，GPT-5.6 Luna 平均高出 99.72 分。
单项差距最大的 benchmark：Creative Writing — Gemini 3.5 Flash-Lite 1,559，GPT-5.6 Luna 1,829（分差 -269.60）。
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。