DataLearner 标志

Gemini 3.1 Pro PreviewvsClaude Opus 4.6

在 19 个共同 benchmark 中，Gemini 3.1 Pro Preview 整体领先：Gemini 3.1 Pro Preview 领先 12 项，Claude Opus 4.6 领先 6 项，持平 1 项，平均分差 +2.43。

Google Deep Mind
Gemini 3.1 Pro Preview

Google Deep Mind · 2026-02-20 · 多模态大模型

Anthropic
Claude Opus 4.6

Anthropic · 2026-02-05 · 推理大模型

Gemini 3.1 Pro Preview12 项(63%)持平1(32%)6 项Claude Opus 4.6

评测分数

按能力类目分组，每组内按分差大小排列；共 19 项。

General Knowledge

Gemini 3.1 Pro Preview 领先 4/6
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
ARC-AGI-277.109 / 62Thinking High (No Tools)66.3017 / 62Extended (no tools)+10.80
LiveBench79.933 / 115Thinking High (No Tools)76.338 / 115Thinking High (No Tools)+3.60
GPQA Diamond94.303 / 187Thinking High (No Tools)91.3115 / 187Extended (no tools)+2.99
HLE51.4022 / 172Thinking High (With Tools)5318 / 172Extended (with tools, internet)-1.60
MMLU92.603 / 66Thinking High (No Tools)91.057 / 66Extended (no tools)+1.55
ARC-AGI-306 / 9Thinking High (No Tools)04 / 9最高（无工具）持平

AI Agent - Tool Usage

Gemini 3.1 Pro Preview 领先 3/3
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
OSWorld-Verified76.2011 / 24Thinking (With Tools)72.7015 / 24Extended (with tools)+3.50
Terminal Bench 2.068.508 / 47Thinking High (With Tools)65.4011 / 47Extended (with tools)+3.10
MCP-Atlas78.209 / 27Thinking High (With Tools)76.8010 / 27Deep Thinking (With Tools)+1.40

Agent Level Benchmark

胶着 2/2
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
τ²-Bench90.802 / 43Thinking High (With Tools)91.891 / 43Extended (with tools)-1.09
τ²-Bench - Telecom99.301 / 35Thinking High (With Tools)99.252 / 35Extended (with tools)+0.05

Coding and Software Engineer

胶着 2/2
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
LiveCodeBench91.703 / 123Thinking High (With Tools)7638 / 123Extended (no tools)+15.70
SWE-bench Verified80.6011 / 112Thinking High (With Tools)80.8410 / 112Extended (with tools)-0.24

Math and Reasoning

Claude Opus 4.6 领先 2/2
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
FrontierMath - Tier 416.7020 / 80Normal (No Tools)22.9012 / 80最高（无工具）-6.20
FrontierMath36.9011 / 60Thinking High (No Tools)40.707 / 60最高（无工具）-3.80

AI Agent - Information Search

Gemini 3.1 Pro Preview 领先 1/1
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
BrowseComp85.905 / 53Thinking High (With Tools + Internet)8411 / 53Thinking (With Tools + Internet)+1.90

Claw-style Agent Evaluation

Claude Opus 4.6 领先 1/1
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
Pinch Bench86.7010 / 37Thinking (With Tools)87.407 / 37Thinking (With Tools)-0.70

Multimodal Understanding

Gemini 3.1 Pro Preview 领先 1/1
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
MMMU80.5012 / 29Thinking High (No Tools)77.3016 / 29Extended (with tools)+3.20

常识推理

Gemini 3.1 Pro Preview 领先 1/1
评测项Gemini 3.1 Pro PreviewClaude Opus 4.6分差
Simple Bench79.602 / 63Normal (No Tools)67.608 / 63Normal (No Tools)+12

规格对比

字段Gemini 3.1 Pro PreviewClaude Opus 4.6
发布机构Google Deep MindAnthropic
发布时间2026-02-202026-02-05
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1M1000K
最大输出64K64K

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 3.1 Pro PreviewClaude Opus 4.6
文本输入$2 / 1M tokens$0.5 / 1M tokens
文本输出$12 / 1M tokens$25 / 1M tokens
缓存读取暂无公开价格$0.5 / 1M tokens
缓存写入暂无公开价格$10 / 1M tokens

小结

  • Gemini 3.1 Pro Preview在以下类目领先:General Knowledge (4/6)、AI Agent - Tool Usage (3/3)、AI Agent - Information Search (1/1)、Multimodal Understanding (1/1)、常识推理 (1/1)
  • Claude Opus 4.6在以下类目领先:Math and Reasoning (2/2)、Claw-style Agent Evaluation (1/1)
  • 胶着类目:Agent Level Benchmark、Coding and Software Engineer

19 个共同 benchmark 上，Gemini 3.1 Pro Preview 平均高出 2.43 分。

单项差距最大的 benchmark：LiveCodeBench — Gemini 3.1 Pro Preview 91.70，Claude Opus 4.6 76（分差 +15.70）。

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。

Gemini 3.1 Pro Preview 详情Claude Opus 4.6 详情·在工具里自定义对比