Tech **[EN] Google Gemini 3 Performance & Usage Guide** **[KO] 구글 제미나이 3 성능·활용 완전 정리*…
Page Info
Writer AndyKim
Hit 1,456 Hit
Date 25-11-27 11:39
Content
**[EN] Google Gemini 3 Performance & Usage Guide**
**[KO] 구글 제미나이 3 성능·활용 완전 정리**
---
## 1. English – What is Gemini 3 and How Strong Is It?
### 1) Overview
**Gemini 3** is Google’s newest and most intelligent Gemini model family (released November 2025). It’s positioned as:
* **“Our most intelligent model yet”** with state-of-the-art reasoning. ([blog.google][1])
* **Best in the world for multimodal understanding** in Google’s own description. ([Google AI for Developers][2])
* The core brain behind the Gemini app and now integrated directly into **Google Search (AI Mode)** and across Google products. ([Reuters][3])
Main variants:
* **Gemini 3 Pro** – flagship LLM for complex reasoning, coding, agents. ([Google AI for Developers][4])
* **Gemini 3 Pro Image** – paired image model (underpins **Nano Banana Pro** for high-end image generation & editing). ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – heavier, slower mode (coming soon) for maximum depth of reasoning. ([Google DeepMind][5])
---
### 2) Core Technical Specs (Gemini 3 Pro Preview)
From the official model card: ([Google AI for Developers][2])
* **Inputs:** text, images, video, audio, PDFs
* **Output:** text
* **Context window:**
* **Input:** up to **1,048,576 tokens (~1M tokens)**
* **Output:** up to **65,536 tokens**
* **Key capabilities:**
* Long-context reasoning over books, codebases, hours of video/audio
* **Function calling**, **code execution**, **file search**, **structured outputs**, **search grounding** (with Google Search)
* “Thinking” mode with adjustable depth (dynamic vs low) for more or less deliberation. ([Google AI for Developers][4])
* **Knowledge cutoff:** **January 2025** for the preview model. ([Google AI for Developers][2])
This means it can:
* Read **hundreds of thousands of words at once**, or multiple long PDFs, logs, or transcripts.
* Call tools and APIs to browse, run code, or work with external systems (e.g., via Gemini API / Google Antigravity). ([Google AI for Developers][4])
---
### 3) Benchmark Performance – How Strong Is It Compared to Others?
From Google DeepMind’s official performance table, Gemini 3 Pro is **state-of-the-art across many benchmarks**, often outperforming earlier Gemini and other frontier models. ([Google DeepMind][5])
Some key highlights (all numbers are for **Gemini 3 Pro** unless noted):
1. **Academic & logical reasoning**
* **Humanity’s Last Exam (reasoning + knowledge, no tools):**
* Gemini 3 Pro: **37.5%**
* Gemini 2.5 Pro: 21.6%
* Claude Sonnet 4.5: 13.7%
* GPT-5.1: 26.5% ([Google DeepMind][5])
* With search & code execution: **45.8%** for Gemini 3 Pro (others not reported). ([Google DeepMind][5])
2. **Visual & multimodal reasoning**
* **ARC-AGI-2 (hard visual reasoning puzzles):**
* G3 Pro: **31.1%** vs 2.5 Pro: 4.9%, Claude Sonnet 4.5: 13.6%, GPT-5.1: 17.6%. ([Google DeepMind][5])
* **MMMU-Pro (multimodal university-level tasks):**
* G3 Pro: **81.0%** vs 2.5 Pro: 68%, GPT-5.1: 76%. ([Google DeepMind][5])
* **ScreenSpot-Pro (UI/screen understanding):**
* G3 Pro: **72.7%**, vastly above 2.5 Pro (11.4%), Claude (36.2%), GPT-5.1 (3.5%). ([Google DeepMind][5])
3. **Science & math**
* **GPQA Diamond (hard scientific QA):** 91.9% (2.5 Pro: 86.4%, GPT-5.1: 88.1%). ([Google DeepMind][5])
* **AIME 2025 math (no tools):** 95% (2.5 Pro: 88%, GPT-5.1: 94%). With code execution, Gemini 3 hits **100%** on AIME. ([Google DeepMind][5])
* **MathArena Apex (competition-level math):** 23.4% vs 0.5–1.6% for 2.5 Pro / Claude / GPT-5.1. ([Google DeepMind][5])
4. **Coding & agents**
* **LiveCodeBench Pro (competitive coding Elo):**
* G3 Pro: **2439** vs 2.5 Pro: 1775, Claude: 1418, GPT-5.1: 2243. ([Google DeepMind][5])
* **Terminal-Bench 2.0 (agentic terminal coding):** 54.2% vs 32.6% (2.5 Pro) & 47.6% (GPT-5.1). ([Google DeepMind][5])
* **SWE-Bench Verified (single attempt agentic coding):** 76.2% (similar to GPT-5.1 and Claude Sonnet 4.5). ([Google DeepMind][5])
* **τ2-bench (tool use):** 85.4% vs 54.9% (2.5 Pro) & 80.2% (GPT-5.1). ([Google DeepMind][5])
5. **Long-context & knowledge**
* **MRCR v2 (8-needle retrieval, 128k context):** 77.0% vs 58.0% (2.5 Pro) & 61.6% (GPT-5.1).
* **1M-token test:** 26.3% vs 16.4% (2.5 Pro); other models in this table don’t support 1M context. ([Google DeepMind][5])
* **FACTS benchmark suite (internal grounding/search/parametric/MM):** 70.5% vs 63.4% (2.5 Pro) & ~50% range for others. ([Google DeepMind][5])
6. **Multilingual & common-sense**
* **MMMLU (multilingual QA):** 91.8% (2.5 Pro: 89.5%; GPT-5.1: 91.0%).
* **Global PIQA (commonsense across 100 languages):** 93.4% vs ~91% for other frontier models. ([Google DeepMind][5])
External write-ups generally place Gemini 3 as one of the very top frontier models alongside GPT-5.1, Claude 4.5, etc. ([Fello AI][6])
---
### 4) Real-World Strengths (From Early Partners)
DeepMind showcases feedback from partners like GitHub, JetBrains, Figma, Shopify, Thomson Reuters, Rakuten, Wayfair, etc.: ([Google DeepMind][5])
* **GitHub Copilot:** ~35% higher accuracy on software engineering challenges vs Gemini 2.5 Pro.
* **JetBrains:** >50% improvement over 2.5 Pro in the number of solved internal coding benchmarks.
* **Rakuten:** >50% better than baselines for extracting structured data from poor-quality document photos, and strong transcription of 3-hour multilingual meetings.
* **Shopify, Manus, Thomson Reuters, Wayfair** report big gains in tool-calling reliability, legal reasoning, and structured business workflows.
These are vendor-supplied numbers but consistent with the benchmark table: Gemini 3 makes a **big jump in reasoning, coding, multimodal understanding, and tool-use reliability**. ([Google DeepMind][5])
---
### 5) Typical Use Cases & Tips (for Users and Developers)
**Great tasks for Gemini 3:**
1. **Complex planning & reasoning**
* Long travel itineraries with constraints, business planning, system design, multi-step workflows. ([Google DeepMind][5])
2. **Serious coding & refactoring**
* Whole-repo analysis, multi-file refactors, generating thousands of lines of UI / frontend code, debugging long logs. Partners like Cursor, Cline, JetBrains, GitHub all emphasize this use case. ([Google DeepMind][5])
3. **Multimodal understanding**
* Reading PDFs with charts, screenshots, forms, or mixing text + images + video (via Gemini 3 Pro + associated media models like Nano Banana Pro & Veo 3.1). ([TechRadar][7])
4. **Agents and automation**
* “Gemini 3 Agent” style tools (Project Mariner / Gemini Agent) that can use Search, tools, and APIs to execute tasks (e.g., research + draft + summarize + format). ([Android Central][8])
5. **Design & UI**
* Vibe-coding frontends, translating wireframes to production code, generating prototypes in tools like Figma, Replit, and Antigravity. ([Google DeepMind][5])
**Usage tips:**
* **Use “dynamic thinking” for hard problems, “low thinking” for speed.**
Gemini 3 Pro defaults to **dynamic thinking**; you can constrain to **low** for lower latency when full reasoning isn’t needed. ([Google AI for Developers][4])
* **Exploit long context.**
Instead of chopping everything into tiny prompts, pass big chunks (docs, code, logs) and let the model cross-reference inside the 1M-token window. ([Google AI for Developers][2])
* **Combine with tools.**
Use search grounding + code execution when doing math, research, or coding tasks; benchmarks show large jumps when tools are enabled. ([Google DeepMind][5])
---
### 6) Limitations & Precautions
Even though Gemini 3 is extremely strong, it is still a generative model:
* **Hallucinations:**
It can still invent citations or misinterpret ambiguous prompts. Cross-check critical facts (especially in law, medicine, finance). ([Google DeepMind][5])
* **Safety & robustness:**
Google applies the Frontier Safety Framework and safety councils, but external security testing for prior models (e.g., 2.0 Flash) still found critical and high-severity issues that had to be mitigated. ([modelcards.withgoogle.com][9])
* **Latency & cost:**
Deep, high-thinking, 1M-token prompts are expensive and slower. Use shorter prompts, low thinking, or a lighter model (e.g., 2.5 Flash / Flash-Lite) for simple tasks. ([Google AI for Developers][2])
* **Not an oracle:**
Benchmarks are curated tests; real-world data is messier. Always add your own validation, logging, and guardrails for production systems. ([Google DeepMind][5])
---
## 2. 한국어 – 제미나이 3 성능·특징 정리
### 1) 개요
* **제미나이 3(Gemini 3)** 는 2025년 11월 기준 구글의 **최신·최상위 대형 모델 패밀리**입니다. ([blog.google][1])
* 구글은 이를 **“지금까지 만든 것 중 가장 똑똑한 AI 모델”**, **“멀티모달 이해에서 세계 최고 수준”**이라고 설명합니다. ([Google AI for Developers][2])
* 제미나이 앱, AI Studio, Vertex AI, 그리고 Google 검색 AI 모드까지 전반에 들어가 있는 핵심 엔진입니다. ([Reuters][3])
주요 라인업:
* **Gemini 3 Pro** – 복잡한 추론·코딩·에이전트에 최적화된 기본 플래그십
* **Gemini 3 Pro Image** – 이미지 생성·편집 전용(여기에 기반한 이미지 모델이 **Nano Banana Pro**) ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – 더 느리지만 추론 성능을 극대화한 모드(추가 공개 예정). ([Google DeepMind][5])
---
### 2) 스펙 요약 (Gemini 3 Pro Preview 기준)
공식 문서 기준: ([Google AI for Developers][2])
* **입력(Input)**: 텍스트, 이미지, 비디오, 오디오, PDF
* **출력(Output)**: 텍스트
* **컨텍스트 길이**:
* 입력 최대 **1,048,576 토큰(약 1M)**
* 출력 최대 **65,536 토큰**
* **기능**: 함수 호출, 코드 실행, 파일 검색, 구조화 출력(JSON 등), Google Search 연동(검색 그라운딩), 긴 문맥 처리 등
* **“생각(Thinking)” 모드 지원**: dynamic(깊게 생각) / low(빠른 응답) 모드 선택 가능. ([Google AI for Developers][4])
* **지식 컷오프**: 2025년 1월(프리뷰 기준). ([Google AI for Developers][2])
실제 의미:
* 책 여러 권 분량, 코드 리포 전체, 긴 로그·회의록·동영상을 한 번에 넣고 **교차 참조·요약·분석** 가능.
* API/도구를 호출해 검색·코드 실행·외부 시스템 조작까지 가능한 **에이전트형 LLM** 역할 수행. ([Google DeepMind][5])
---
### 3) 벤치마크 성능 (간단 비교)
DeepMind 공식 성능 표 기준: ([Google DeepMind][5])
* **Humanity’s Last Exam (고난도 종합 추론)**
* 3 Pro: **37.5%**, 2.5 Pro: 21.6%, GPT-5.1: 26.5%
* **ARC-AGI-2 (추상·시각 추론 퍼즐)**
* 3 Pro: **31.1%**, 2.5 Pro: 4.9%, GPT-5.1: 17.6%
* **GPQA Diamond (과학 지식)**
* 3 Pro: 91.9%, 2.5 Pro: 86.4%, GPT-5.1: 88.1%
* **AIME 2025 (수학)**
* 도구 없이: 3 Pro 95%, GPT-5.1 94%
* 코드 실행 사용 시: 3 Pro 100%
* **MMMU-Pro (멀티모달 대학 수준 문제)**
* 3 Pro: **81%**, 2.5 Pro: 68%, GPT-5.1: 76%
* **ScreenSpot-Pro (UI/스크린 이해)**
* 3 Pro: **72.7%**, 2.5 Pro: 11.4%, GPT-5.1: 3.5%
* **LiveCodeBench Pro (코딩 Elo)**: 3 Pro 2439, 2.5 Pro 1775, GPT-5.1 2243
* **길게·멀리 보는 에이전트 작업(Vending-Bench 등)**에서도 2.5 Pro 및 타 모델 대비 큰 격차를 보입니다. ([Google DeepMind][5])
요약하면, **추론·수학·시각·멀티모달·코딩·에이전트·롱컨텍스트** 거의 모든 영역에서 2.5 Pro보다 크게 앞서고, 다른 최상위 모델들과 비교해도 상위권 또는 1위를 차지하는 항목이 많습니다. ([Google DeepMind][5])
---
### 4) 실제 사용에서 강점
* **코딩**: GitHub, JetBrains, Cursor 등의 피드백에 따르면 2.5 Pro 대비 35~50% 이상 성능 향상(문제 해결율 기준) 보고. ([Google DeepMind][5])
* **문서·비즈니스 워크플로**: Rakuten, Wayfair 등에서는 긴 회의록·문서에서 구조화된 데이터 추출, 인포그래픽 생성 등에서 “기존 모델 대비 50% 이상 향상” 같은 숫자를 제시. ([Google DeepMind][5])
* **디자인·UI**: Figma·Replit 등은 프론트엔드·디자인 쪽에서 다양한 스타일·레이아웃·인터랙션을 안정적으로 생성한다고 평가. ([Google DeepMind][5])
---
### 5) 활용 팁 (개발자·일반 사용자 공통)
1. **난도에 따라 Thinking 모드 조절**
* 복잡한 설계·알고리즘·법률 문서 해석 등에는 **dynamic(깊은 추론)** 모드.
* 단순 요약·포맷 변환·짧은 답변은 **low(빠른)** 모드로 비용·지연을 줄이기. ([Google AI for Developers][4])
2. **롱컨텍스트를 적극 활용**
* PPT, PDF, 스펙 문서, 전체 Git 리포, 긴 로그를 **그대로** 넣고 “전체 구조 분석 + 세부 버그 찾기 + 요약” 과 같이 한 번에 시키는 편이 효율적입니다. ([Google AI for Developers][2])
3. **검색·코드 실행과 같이 사용**
* 수학·과학·코드·최신 정보가 필요한 작업에서는 Search grounding, code execution, 도구 호출을 켜두면 벤치마크에서도 성능이 눈에 띄게 올라갑니다. ([Google DeepMind][5])
4. **에이전트 패턴**
* “문제 분석 → 필요한 자료 검색 → 코드 작성·실행 → 결과 정리 → 리포트 생성” 같은 **멀티스텝 워크플로우**를 한 번에 정의하고 Gemini 3에 맡기는 방식이 잘 맞습니다. (Project Mariner / Gemini Agent / Antigravity 연동 등) ([Android Central][8])
---
### 6) 주의해야 할 점
* **할루시네이션**: 숫자·법률·의료·재무 정보는 반드시 2차 검증 필요.
* **보안·안전**: 구글은 Frontier Safety Framework, RSC 검토 등 안전 프로세스를 적용하지만, 2.0 Flash에서도 보안 테스트에서 여러 등급의 취약점이 발견된 전례가 있습니다. 3에서도 **권한·도구 사용 범위·출력 검증 레이어**를 직접 두는 것이 좋습니다. ([modelcards.withgoogle.com][9])
* **비용·속도**: 1M 토큰 + dynamic thinking은 매우 비쌀 수 있습니다. 단순 작업에는 2.5 Flash / Flash-Lite 같은 경량 모델을 쓰는 구조가 일반적입니다. ([Google DeepMind][5])
---
## 3. 日本語 – Gemini 3 の性能と活用ポイント
### 1) 概要
* **Gemini 3** は Google が 2025 年に公開した最新のフラグシップ LLM で、「これまでで最も高度なモデル」と公式に位置づけられています。 ([blog.google][1])
* Gemini アプリ、Google 検索の AI モード、AI Studio、Vertex AI、さらに Antigravity(新しいエージェント開発環境)などに統合されています。 ([Android Central][8])
バリエーション:
* **Gemini 3 Pro** – 高度な推論・マルチモーダル・コーディング・エージェント向け
* **Gemini 3 Pro Image** – 画像生成・編集 (Nano Banana Pro のベース) ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – さらに深い推論に特化したモード（公開準備中） ([Google DeepMind][5])
---
### 2) 主要スペック（3 Pro）
([Google AI for Developers][2])
* 入力: テキスト / 画像 / 動画 / 音声 / PDF
* 出力: テキスト
* コンテキスト長: 入力最大約 **100 万トークン**, 出力最大 **65K トークン**
* 機能: 関数呼び出し、コード実行、ファイル検索、Structured output、Search grounding、Thinking モードなど
* ナレッジカットオフ: 2025 年 1 月
---
### 3) ベンチマーク性能（かんたん比較）
DeepMind の表を要約すると: ([Google DeepMind][5])
* 高度な学術推論 (Humanity’s Last Exam) では 2.5 Pro や他社モデルより **大幅に高得点**
* ARC-AGI-2, MMMU-Pro, ScreenSpot-Pro などのマルチモーダル・視覚推論ベンチマークで**トップクラス**
* GPQA, AIME, MathArena などの科学・数学ベンチマークでも最高レベル
* LiveCodeBench, SWE-Bench, Terminal-Bench, τ2-bench などのコーディング・エージェント系ベンチマークでも 2.5 Pro を大きく上回る
総合的には、**推論・マルチモーダル・コーディング・エージェント・長文コンテキストの全てで 2.5 Pro から大きくジャンプした世代**と考えてよいです。
---
### 4) 使いどころとコツ
**得意分野:**
* 長文ドキュメントやコードベース全体を読み込んでの分析・要約・設計
* Web 検索＋コード実行を組み合わせたリサーチ、数式・アルゴリズム問題
* UI デザインやフロントエンド実装の自動生成（Figma や Replit との連携例あり） ([Google DeepMind][5])
* Agent（エージェント）として複数ステップのタスク処理、ワークフロー自動化 ([구글 개발자 블로그][10])
**運用のポイント:**
* 複雑な課題 → **dynamic thinking** モード
* 日常的な短いタスク → **low thinking** でレイテンシとコストを削減
* 1M トークンの長文を使うときは、事前にテキストを整理・チャンク化してから投入すると精度が安定しやすい
**注意点:**
* いまだにハルシネーションはあり得るため、法律・医療・金融のような領域では人間による確認が必須。 ([Google DeepMind][5])
* 安全性評価は行われているものの（Frontier Safety Framework 等）、攻撃的プロンプトやツール呼び出し権限などは自分側でガードレールを設計する必要があります。 ([modelcards.withgoogle.com][9])
---
## 4. 中文（简体）– Gemini 3 的性能与使用建议
### 1) 概况
* **Gemini 3** 是谷歌在 2025 年推出的最新一代大型模型家族，被官方称为“目前最智能的 Gemini 模型”。 ([blog.google][1])
* 已经深度集成到 Gemini App、Google 搜索 AI 模式、AI Studio、Vertex AI 以及新的代理开发环境 Antigravity 中。 ([Android Central][8])
主要版本：
* **Gemini 3 Pro** – 主力 LLM，用于复杂推理、编程、代理
* **Gemini 3 Pro Image** – 图像生成和编辑（对外产品名为 Nano Banana Pro） ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – 更慢但推理能力更强的模式（即将开放） ([Google DeepMind][5])
---
### 2) 技术参数（3 Pro）
([Google AI for Developers][2])
* 输入：文本 / 图片 / 视频 / 音频 / PDF
* 输出：文本
* 上下文长度：输入约 **100 万 token**，输出 **6.5 万 token**
* 主要能力：函数调用、代码执行、文件检索、结构化输出、与 Google 搜索结合的“搜索增强”、Thinking 思考模式（动态/低思考）等
* 知识截止时间：2025 年 1 月
这意味着：
* 可以一次性读入多本书、整个代码库、长时间视频/会议记录，并在内部交叉分析。
* 通过 API + 工具调用，实现搜索、抓取、运行代码、访问外部系统的“AI 代理”。 ([Google DeepMind][5])
---
### 3) 性能表现（与 2.5 Pro 和其他模型的比较）
根据 DeepMind 官方性能表： ([Google DeepMind][5])
* 在 **Humanity’s Last Exam、GPQA、AIME、MathArena** 等高难度推理和数学测试中，Gemini 3 Pro 明显优于 Gemini 2.5 Pro，并与 GPT-5.1 等顶级模型处于同一水平甚至更高。
* 在 **ARC-AGI-2、MMMU-Pro、ScreenSpot-Pro、CharXiv Reasoning** 等多模态/视觉推理基准中，经常排在第一。
* 在 **LiveCodeBench、SWE-Bench、Terminal-Bench、τ2-bench** 等编码和 Agent 基准中，大幅领先 2.5 Pro。
* 在 **MRCR v2、FACTS、MMMLU、Global PIQA** 等长上下文、多语言、常识和检索类测试中，同样表现一流。
简单来说：**在推理、编码、多模态、长上下文、工具调用这几个方向，Gemini 3 Pro 相比 2.5 Pro 是一代“飞跃式升级”。**
---
### 4) 合适的应用场景与实践建议
**特别适合：**
* 大型项目：整库代码重构、复杂系统设计、长文档分析、从多个 PDF/网页综合出一份报告
* 多模态任务：图表理解、截图/界面解析、低质量文档 OCR + 结构化抽取、会议记录总结 ([Google DeepMind][5])
* 高级 Agent：自动研究、自动编码、自动运行脚本并形成最终报告或可视化结果 ([구글 개발자 블로그][10])
**提示：**
* 复杂任务：用 **dynamic thinking**，允许模型“多想一会儿”。
* 简单任务：用 **low thinking**，响应更快、成本更低。 ([Google AI for Developers][4])
* 1M token 长上下文要配合良好的结构化输入（标题、分节、编号），方便模型建立内部索引。 ([Google AI for Developers][2])
* 对于编程和数学，开启“代码执行”或结合外部工具，能显著提高正确率。 ([Google DeepMind][5])
**注意事项：**
* 仍然可能产生“看起来像真的”但实际上错误的信息，特别是在专业领域。务必自行核实。 ([Google DeepMind][5])
* 构建生产级系统时，要增加权限控制、输入/输出过滤、日志审计和安全策略，以防 prompt 注入、越权操作等风险。 ([modelcards.withgoogle.com][9])
---
* [Business Insider](https://www.businessinsider.com/google-ceo-sundar-pichai-hopes-gemini-ai-team-gets-sleep-2025-11?utm_source=chatgpt.com)
* [TechRadar](https://www.techradar.com/ai-platforms-assistants/gemini/google-launches-nano-banana-pro-a-massive-leap-in-ai-image-editing-powered-by-gemini-3-pro?utm_source=chatgpt.com)
* [The Times of India](https://timesofindia.indiatimes.com/technology/tech-news/google-launches-gemini-3-pro-image-based-nano-banana-pro-ai-model-all-details/articleshow/125469483.cms?utm_source=chatgpt.com)
* [Android Central](https://www.androidcentral.com/apps-software/ai/googles-november-gemini-drop-adds-gemini-3-nano-banana-pro-and-more?utm_source=chatgpt.com)
[1]: https://blog.google/products/gemini/gemini-3/?utm_source=chatgpt.com "A new era of intelligence with Gemini 3"
[2]: https://ai.google.dev/gemini-api/docs/models "Gemini models | Gemini API | Google AI for Developers"
[3]: https://www.reuters.com/business/media-telecom/google-launches-gemini-3-embeds-ai-model-into-search-immediately-2025-11-18/?utm_source=chatgpt.com "Google launches Gemini 3, embeds AI model into search ..."
[4]: https://ai.google.dev/gemini-api/docs/gemini-3 "Gemini 3 Developer Guide | Gemini API | Google AI for Developers"
[5]: https://deepmind.google/models/gemini/ "
Gemini 3 -
Google DeepMind
"
[6]: https://felloai.com/2025/11/the-best-ai-of-november-2025/?utm_source=chatgpt.com "The Best AI of November 2025: Gemini 3 vs GPT-5.1 ..."
[7]: https://www.techradar.com/ai-platforms-assistants/gemini/google-launches-nano-banana-pro-a-massive-leap-in-ai-image-editing-powered-by-gemini-3-pro?utm_source=chatgpt.com "Google launches Nano Banana Pro, a massive leap in AI image editing powered by Gemini 3 Pro"
[8]: https://www.androidcentral.com/apps-software/ai/googles-november-gemini-drop-adds-gemini-3-nano-banana-pro-and-more?utm_source=chatgpt.com "Everything Google added to the Gemini app in November, from Gemini 3 to Nano Banana Pro"
[9]: https://modelcards.withgoogle.com/assets/documents/gemini-2-flash.pdf?utm_source=chatgpt.com "Gemini 2.0 Flash"
[10]: https://developers.googleblog.com/new-gemini-api-updates-for-gemini-3/?utm_source=chatgpt.com "New Gemini API updates for Gemini 3"
**[KO] 구글 제미나이 3 성능·활용 완전 정리**
---
## 1. English – What is Gemini 3 and How Strong Is It?
### 1) Overview
**Gemini 3** is Google’s newest and most intelligent Gemini model family (released November 2025). It’s positioned as:
* **“Our most intelligent model yet”** with state-of-the-art reasoning. ([blog.google][1])
* **Best in the world for multimodal understanding** in Google’s own description. ([Google AI for Developers][2])
* The core brain behind the Gemini app and now integrated directly into **Google Search (AI Mode)** and across Google products. ([Reuters][3])
Main variants:
* **Gemini 3 Pro** – flagship LLM for complex reasoning, coding, agents. ([Google AI for Developers][4])
* **Gemini 3 Pro Image** – paired image model (underpins **Nano Banana Pro** for high-end image generation & editing). ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – heavier, slower mode (coming soon) for maximum depth of reasoning. ([Google DeepMind][5])
---
### 2) Core Technical Specs (Gemini 3 Pro Preview)
From the official model card: ([Google AI for Developers][2])
* **Inputs:** text, images, video, audio, PDFs
* **Output:** text
* **Context window:**
* **Input:** up to **1,048,576 tokens (~1M tokens)**
* **Output:** up to **65,536 tokens**
* **Key capabilities:**
* Long-context reasoning over books, codebases, hours of video/audio
* **Function calling**, **code execution**, **file search**, **structured outputs**, **search grounding** (with Google Search)
* “Thinking” mode with adjustable depth (dynamic vs low) for more or less deliberation. ([Google AI for Developers][4])
* **Knowledge cutoff:** **January 2025** for the preview model. ([Google AI for Developers][2])
This means it can:
* Read **hundreds of thousands of words at once**, or multiple long PDFs, logs, or transcripts.
* Call tools and APIs to browse, run code, or work with external systems (e.g., via Gemini API / Google Antigravity). ([Google AI for Developers][4])
---
### 3) Benchmark Performance – How Strong Is It Compared to Others?
From Google DeepMind’s official performance table, Gemini 3 Pro is **state-of-the-art across many benchmarks**, often outperforming earlier Gemini and other frontier models. ([Google DeepMind][5])
Some key highlights (all numbers are for **Gemini 3 Pro** unless noted):
1. **Academic & logical reasoning**
* **Humanity’s Last Exam (reasoning + knowledge, no tools):**
* Gemini 3 Pro: **37.5%**
* Gemini 2.5 Pro: 21.6%
* Claude Sonnet 4.5: 13.7%
* GPT-5.1: 26.5% ([Google DeepMind][5])
* With search & code execution: **45.8%** for Gemini 3 Pro (others not reported). ([Google DeepMind][5])
2. **Visual & multimodal reasoning**
* **ARC-AGI-2 (hard visual reasoning puzzles):**
* G3 Pro: **31.1%** vs 2.5 Pro: 4.9%, Claude Sonnet 4.5: 13.6%, GPT-5.1: 17.6%. ([Google DeepMind][5])
* **MMMU-Pro (multimodal university-level tasks):**
* G3 Pro: **81.0%** vs 2.5 Pro: 68%, GPT-5.1: 76%. ([Google DeepMind][5])
* **ScreenSpot-Pro (UI/screen understanding):**
* G3 Pro: **72.7%**, vastly above 2.5 Pro (11.4%), Claude (36.2%), GPT-5.1 (3.5%). ([Google DeepMind][5])
3. **Science & math**
* **GPQA Diamond (hard scientific QA):** 91.9% (2.5 Pro: 86.4%, GPT-5.1: 88.1%). ([Google DeepMind][5])
* **AIME 2025 math (no tools):** 95% (2.5 Pro: 88%, GPT-5.1: 94%). With code execution, Gemini 3 hits **100%** on AIME. ([Google DeepMind][5])
* **MathArena Apex (competition-level math):** 23.4% vs 0.5–1.6% for 2.5 Pro / Claude / GPT-5.1. ([Google DeepMind][5])
4. **Coding & agents**
* **LiveCodeBench Pro (competitive coding Elo):**
* G3 Pro: **2439** vs 2.5 Pro: 1775, Claude: 1418, GPT-5.1: 2243. ([Google DeepMind][5])
* **Terminal-Bench 2.0 (agentic terminal coding):** 54.2% vs 32.6% (2.5 Pro) & 47.6% (GPT-5.1). ([Google DeepMind][5])
* **SWE-Bench Verified (single attempt agentic coding):** 76.2% (similar to GPT-5.1 and Claude Sonnet 4.5). ([Google DeepMind][5])
* **τ2-bench (tool use):** 85.4% vs 54.9% (2.5 Pro) & 80.2% (GPT-5.1). ([Google DeepMind][5])
5. **Long-context & knowledge**
* **MRCR v2 (8-needle retrieval, 128k context):** 77.0% vs 58.0% (2.5 Pro) & 61.6% (GPT-5.1).
* **1M-token test:** 26.3% vs 16.4% (2.5 Pro); other models in this table don’t support 1M context. ([Google DeepMind][5])
* **FACTS benchmark suite (internal grounding/search/parametric/MM):** 70.5% vs 63.4% (2.5 Pro) & ~50% range for others. ([Google DeepMind][5])
6. **Multilingual & common-sense**
* **MMMLU (multilingual QA):** 91.8% (2.5 Pro: 89.5%; GPT-5.1: 91.0%).
* **Global PIQA (commonsense across 100 languages):** 93.4% vs ~91% for other frontier models. ([Google DeepMind][5])
External write-ups generally place Gemini 3 as one of the very top frontier models alongside GPT-5.1, Claude 4.5, etc. ([Fello AI][6])
---
### 4) Real-World Strengths (From Early Partners)
DeepMind showcases feedback from partners like GitHub, JetBrains, Figma, Shopify, Thomson Reuters, Rakuten, Wayfair, etc.: ([Google DeepMind][5])
* **GitHub Copilot:** ~35% higher accuracy on software engineering challenges vs Gemini 2.5 Pro.
* **JetBrains:** >50% improvement over 2.5 Pro in the number of solved internal coding benchmarks.
* **Rakuten:** >50% better than baselines for extracting structured data from poor-quality document photos, and strong transcription of 3-hour multilingual meetings.
* **Shopify, Manus, Thomson Reuters, Wayfair** report big gains in tool-calling reliability, legal reasoning, and structured business workflows.
These are vendor-supplied numbers but consistent with the benchmark table: Gemini 3 makes a **big jump in reasoning, coding, multimodal understanding, and tool-use reliability**. ([Google DeepMind][5])
---
### 5) Typical Use Cases & Tips (for Users and Developers)
**Great tasks for Gemini 3:**
1. **Complex planning & reasoning**
* Long travel itineraries with constraints, business planning, system design, multi-step workflows. ([Google DeepMind][5])
2. **Serious coding & refactoring**
* Whole-repo analysis, multi-file refactors, generating thousands of lines of UI / frontend code, debugging long logs. Partners like Cursor, Cline, JetBrains, GitHub all emphasize this use case. ([Google DeepMind][5])
3. **Multimodal understanding**
* Reading PDFs with charts, screenshots, forms, or mixing text + images + video (via Gemini 3 Pro + associated media models like Nano Banana Pro & Veo 3.1). ([TechRadar][7])
4. **Agents and automation**
* “Gemini 3 Agent” style tools (Project Mariner / Gemini Agent) that can use Search, tools, and APIs to execute tasks (e.g., research + draft + summarize + format). ([Android Central][8])
5. **Design & UI**
* Vibe-coding frontends, translating wireframes to production code, generating prototypes in tools like Figma, Replit, and Antigravity. ([Google DeepMind][5])
**Usage tips:**
* **Use “dynamic thinking” for hard problems, “low thinking” for speed.**
Gemini 3 Pro defaults to **dynamic thinking**; you can constrain to **low** for lower latency when full reasoning isn’t needed. ([Google AI for Developers][4])
* **Exploit long context.**
Instead of chopping everything into tiny prompts, pass big chunks (docs, code, logs) and let the model cross-reference inside the 1M-token window. ([Google AI for Developers][2])
* **Combine with tools.**
Use search grounding + code execution when doing math, research, or coding tasks; benchmarks show large jumps when tools are enabled. ([Google DeepMind][5])
---
### 6) Limitations & Precautions
Even though Gemini 3 is extremely strong, it is still a generative model:
* **Hallucinations:**
It can still invent citations or misinterpret ambiguous prompts. Cross-check critical facts (especially in law, medicine, finance). ([Google DeepMind][5])
* **Safety & robustness:**
Google applies the Frontier Safety Framework and safety councils, but external security testing for prior models (e.g., 2.0 Flash) still found critical and high-severity issues that had to be mitigated. ([modelcards.withgoogle.com][9])
* **Latency & cost:**
Deep, high-thinking, 1M-token prompts are expensive and slower. Use shorter prompts, low thinking, or a lighter model (e.g., 2.5 Flash / Flash-Lite) for simple tasks. ([Google AI for Developers][2])
* **Not an oracle:**
Benchmarks are curated tests; real-world data is messier. Always add your own validation, logging, and guardrails for production systems. ([Google DeepMind][5])
---
## 2. 한국어 – 제미나이 3 성능·특징 정리
### 1) 개요
* **제미나이 3(Gemini 3)** 는 2025년 11월 기준 구글의 **최신·최상위 대형 모델 패밀리**입니다. ([blog.google][1])
* 구글은 이를 **“지금까지 만든 것 중 가장 똑똑한 AI 모델”**, **“멀티모달 이해에서 세계 최고 수준”**이라고 설명합니다. ([Google AI for Developers][2])
* 제미나이 앱, AI Studio, Vertex AI, 그리고 Google 검색 AI 모드까지 전반에 들어가 있는 핵심 엔진입니다. ([Reuters][3])
주요 라인업:
* **Gemini 3 Pro** – 복잡한 추론·코딩·에이전트에 최적화된 기본 플래그십
* **Gemini 3 Pro Image** – 이미지 생성·편집 전용(여기에 기반한 이미지 모델이 **Nano Banana Pro**) ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – 더 느리지만 추론 성능을 극대화한 모드(추가 공개 예정). ([Google DeepMind][5])
---
### 2) 스펙 요약 (Gemini 3 Pro Preview 기준)
공식 문서 기준: ([Google AI for Developers][2])
* **입력(Input)**: 텍스트, 이미지, 비디오, 오디오, PDF
* **출력(Output)**: 텍스트
* **컨텍스트 길이**:
* 입력 최대 **1,048,576 토큰(약 1M)**
* 출력 최대 **65,536 토큰**
* **기능**: 함수 호출, 코드 실행, 파일 검색, 구조화 출력(JSON 등), Google Search 연동(검색 그라운딩), 긴 문맥 처리 등
* **“생각(Thinking)” 모드 지원**: dynamic(깊게 생각) / low(빠른 응답) 모드 선택 가능. ([Google AI for Developers][4])
* **지식 컷오프**: 2025년 1월(프리뷰 기준). ([Google AI for Developers][2])
실제 의미:
* 책 여러 권 분량, 코드 리포 전체, 긴 로그·회의록·동영상을 한 번에 넣고 **교차 참조·요약·분석** 가능.
* API/도구를 호출해 검색·코드 실행·외부 시스템 조작까지 가능한 **에이전트형 LLM** 역할 수행. ([Google DeepMind][5])
---
### 3) 벤치마크 성능 (간단 비교)
DeepMind 공식 성능 표 기준: ([Google DeepMind][5])
* **Humanity’s Last Exam (고난도 종합 추론)**
* 3 Pro: **37.5%**, 2.5 Pro: 21.6%, GPT-5.1: 26.5%
* **ARC-AGI-2 (추상·시각 추론 퍼즐)**
* 3 Pro: **31.1%**, 2.5 Pro: 4.9%, GPT-5.1: 17.6%
* **GPQA Diamond (과학 지식)**
* 3 Pro: 91.9%, 2.5 Pro: 86.4%, GPT-5.1: 88.1%
* **AIME 2025 (수학)**
* 도구 없이: 3 Pro 95%, GPT-5.1 94%
* 코드 실행 사용 시: 3 Pro 100%
* **MMMU-Pro (멀티모달 대학 수준 문제)**
* 3 Pro: **81%**, 2.5 Pro: 68%, GPT-5.1: 76%
* **ScreenSpot-Pro (UI/스크린 이해)**
* 3 Pro: **72.7%**, 2.5 Pro: 11.4%, GPT-5.1: 3.5%
* **LiveCodeBench Pro (코딩 Elo)**: 3 Pro 2439, 2.5 Pro 1775, GPT-5.1 2243
* **길게·멀리 보는 에이전트 작업(Vending-Bench 등)**에서도 2.5 Pro 및 타 모델 대비 큰 격차를 보입니다. ([Google DeepMind][5])
요약하면, **추론·수학·시각·멀티모달·코딩·에이전트·롱컨텍스트** 거의 모든 영역에서 2.5 Pro보다 크게 앞서고, 다른 최상위 모델들과 비교해도 상위권 또는 1위를 차지하는 항목이 많습니다. ([Google DeepMind][5])
---
### 4) 실제 사용에서 강점
* **코딩**: GitHub, JetBrains, Cursor 등의 피드백에 따르면 2.5 Pro 대비 35~50% 이상 성능 향상(문제 해결율 기준) 보고. ([Google DeepMind][5])
* **문서·비즈니스 워크플로**: Rakuten, Wayfair 등에서는 긴 회의록·문서에서 구조화된 데이터 추출, 인포그래픽 생성 등에서 “기존 모델 대비 50% 이상 향상” 같은 숫자를 제시. ([Google DeepMind][5])
* **디자인·UI**: Figma·Replit 등은 프론트엔드·디자인 쪽에서 다양한 스타일·레이아웃·인터랙션을 안정적으로 생성한다고 평가. ([Google DeepMind][5])
---
### 5) 활용 팁 (개발자·일반 사용자 공통)
1. **난도에 따라 Thinking 모드 조절**
* 복잡한 설계·알고리즘·법률 문서 해석 등에는 **dynamic(깊은 추론)** 모드.
* 단순 요약·포맷 변환·짧은 답변은 **low(빠른)** 모드로 비용·지연을 줄이기. ([Google AI for Developers][4])
2. **롱컨텍스트를 적극 활용**
* PPT, PDF, 스펙 문서, 전체 Git 리포, 긴 로그를 **그대로** 넣고 “전체 구조 분석 + 세부 버그 찾기 + 요약” 과 같이 한 번에 시키는 편이 효율적입니다. ([Google AI for Developers][2])
3. **검색·코드 실행과 같이 사용**
* 수학·과학·코드·최신 정보가 필요한 작업에서는 Search grounding, code execution, 도구 호출을 켜두면 벤치마크에서도 성능이 눈에 띄게 올라갑니다. ([Google DeepMind][5])
4. **에이전트 패턴**
* “문제 분석 → 필요한 자료 검색 → 코드 작성·실행 → 결과 정리 → 리포트 생성” 같은 **멀티스텝 워크플로우**를 한 번에 정의하고 Gemini 3에 맡기는 방식이 잘 맞습니다. (Project Mariner / Gemini Agent / Antigravity 연동 등) ([Android Central][8])
---
### 6) 주의해야 할 점
* **할루시네이션**: 숫자·법률·의료·재무 정보는 반드시 2차 검증 필요.
* **보안·안전**: 구글은 Frontier Safety Framework, RSC 검토 등 안전 프로세스를 적용하지만, 2.0 Flash에서도 보안 테스트에서 여러 등급의 취약점이 발견된 전례가 있습니다. 3에서도 **권한·도구 사용 범위·출력 검증 레이어**를 직접 두는 것이 좋습니다. ([modelcards.withgoogle.com][9])
* **비용·속도**: 1M 토큰 + dynamic thinking은 매우 비쌀 수 있습니다. 단순 작업에는 2.5 Flash / Flash-Lite 같은 경량 모델을 쓰는 구조가 일반적입니다. ([Google DeepMind][5])
---
## 3. 日本語 – Gemini 3 の性能と活用ポイント
### 1) 概要
* **Gemini 3** は Google が 2025 年に公開した最新のフラグシップ LLM で、「これまでで最も高度なモデル」と公式に位置づけられています。 ([blog.google][1])
* Gemini アプリ、Google 検索の AI モード、AI Studio、Vertex AI、さらに Antigravity(新しいエージェント開発環境)などに統合されています。 ([Android Central][8])
バリエーション:
* **Gemini 3 Pro** – 高度な推論・マルチモーダル・コーディング・エージェント向け
* **Gemini 3 Pro Image** – 画像生成・編集 (Nano Banana Pro のベース) ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – さらに深い推論に特化したモード（公開準備中） ([Google DeepMind][5])
---
### 2) 主要スペック（3 Pro）
([Google AI for Developers][2])
* 入力: テキスト / 画像 / 動画 / 音声 / PDF
* 出力: テキスト
* コンテキスト長: 入力最大約 **100 万トークン**, 出力最大 **65K トークン**
* 機能: 関数呼び出し、コード実行、ファイル検索、Structured output、Search grounding、Thinking モードなど
* ナレッジカットオフ: 2025 年 1 月
---
### 3) ベンチマーク性能（かんたん比較）
DeepMind の表を要約すると: ([Google DeepMind][5])
* 高度な学術推論 (Humanity’s Last Exam) では 2.5 Pro や他社モデルより **大幅に高得点**
* ARC-AGI-2, MMMU-Pro, ScreenSpot-Pro などのマルチモーダル・視覚推論ベンチマークで**トップクラス**
* GPQA, AIME, MathArena などの科学・数学ベンチマークでも最高レベル
* LiveCodeBench, SWE-Bench, Terminal-Bench, τ2-bench などのコーディング・エージェント系ベンチマークでも 2.5 Pro を大きく上回る
総合的には、**推論・マルチモーダル・コーディング・エージェント・長文コンテキストの全てで 2.5 Pro から大きくジャンプした世代**と考えてよいです。
---
### 4) 使いどころとコツ
**得意分野:**
* 長文ドキュメントやコードベース全体を読み込んでの分析・要約・設計
* Web 検索＋コード実行を組み合わせたリサーチ、数式・アルゴリズム問題
* UI デザインやフロントエンド実装の自動生成（Figma や Replit との連携例あり） ([Google DeepMind][5])
* Agent（エージェント）として複数ステップのタスク処理、ワークフロー自動化 ([구글 개발자 블로그][10])
**運用のポイント:**
* 複雑な課題 → **dynamic thinking** モード
* 日常的な短いタスク → **low thinking** でレイテンシとコストを削減
* 1M トークンの長文を使うときは、事前にテキストを整理・チャンク化してから投入すると精度が安定しやすい
**注意点:**
* いまだにハルシネーションはあり得るため、法律・医療・金融のような領域では人間による確認が必須。 ([Google DeepMind][5])
* 安全性評価は行われているものの（Frontier Safety Framework 等）、攻撃的プロンプトやツール呼び出し権限などは自分側でガードレールを設計する必要があります。 ([modelcards.withgoogle.com][9])
---
## 4. 中文（简体）– Gemini 3 的性能与使用建议
### 1) 概况
* **Gemini 3** 是谷歌在 2025 年推出的最新一代大型模型家族，被官方称为“目前最智能的 Gemini 模型”。 ([blog.google][1])
* 已经深度集成到 Gemini App、Google 搜索 AI 模式、AI Studio、Vertex AI 以及新的代理开发环境 Antigravity 中。 ([Android Central][8])
主要版本：
* **Gemini 3 Pro** – 主力 LLM，用于复杂推理、编程、代理
* **Gemini 3 Pro Image** – 图像生成和编辑（对外产品名为 Nano Banana Pro） ([Google AI for Developers][2])
* **Gemini 3 Deep Think** – 更慢但推理能力更强的模式（即将开放） ([Google DeepMind][5])
---
### 2) 技术参数（3 Pro）
([Google AI for Developers][2])
* 输入：文本 / 图片 / 视频 / 音频 / PDF
* 输出：文本
* 上下文长度：输入约 **100 万 token**，输出 **6.5 万 token**
* 主要能力：函数调用、代码执行、文件检索、结构化输出、与 Google 搜索结合的“搜索增强”、Thinking 思考模式（动态/低思考）等
* 知识截止时间：2025 年 1 月
这意味着：
* 可以一次性读入多本书、整个代码库、长时间视频/会议记录，并在内部交叉分析。
* 通过 API + 工具调用，实现搜索、抓取、运行代码、访问外部系统的“AI 代理”。 ([Google DeepMind][5])
---
### 3) 性能表现（与 2.5 Pro 和其他模型的比较）
根据 DeepMind 官方性能表： ([Google DeepMind][5])
* 在 **Humanity’s Last Exam、GPQA、AIME、MathArena** 等高难度推理和数学测试中，Gemini 3 Pro 明显优于 Gemini 2.5 Pro，并与 GPT-5.1 等顶级模型处于同一水平甚至更高。
* 在 **ARC-AGI-2、MMMU-Pro、ScreenSpot-Pro、CharXiv Reasoning** 等多模态/视觉推理基准中，经常排在第一。
* 在 **LiveCodeBench、SWE-Bench、Terminal-Bench、τ2-bench** 等编码和 Agent 基准中，大幅领先 2.5 Pro。
* 在 **MRCR v2、FACTS、MMMLU、Global PIQA** 等长上下文、多语言、常识和检索类测试中，同样表现一流。
简单来说：**在推理、编码、多模态、长上下文、工具调用这几个方向，Gemini 3 Pro 相比 2.5 Pro 是一代“飞跃式升级”。**
---
### 4) 合适的应用场景与实践建议
**特别适合：**
* 大型项目：整库代码重构、复杂系统设计、长文档分析、从多个 PDF/网页综合出一份报告
* 多模态任务：图表理解、截图/界面解析、低质量文档 OCR + 结构化抽取、会议记录总结 ([Google DeepMind][5])
* 高级 Agent：自动研究、自动编码、自动运行脚本并形成最终报告或可视化结果 ([구글 개발자 블로그][10])
**提示：**
* 复杂任务：用 **dynamic thinking**，允许模型“多想一会儿”。
* 简单任务：用 **low thinking**，响应更快、成本更低。 ([Google AI for Developers][4])
* 1M token 长上下文要配合良好的结构化输入（标题、分节、编号），方便模型建立内部索引。 ([Google AI for Developers][2])
* 对于编程和数学，开启“代码执行”或结合外部工具，能显著提高正确率。 ([Google DeepMind][5])
**注意事项：**
* 仍然可能产生“看起来像真的”但实际上错误的信息，特别是在专业领域。务必自行核实。 ([Google DeepMind][5])
* 构建生产级系统时，要增加权限控制、输入/输出过滤、日志审计和安全策略，以防 prompt 注入、越权操作等风险。 ([modelcards.withgoogle.com][9])
---
* [Business Insider](https://www.businessinsider.com/google-ceo-sundar-pichai-hopes-gemini-ai-team-gets-sleep-2025-11?utm_source=chatgpt.com)
* [TechRadar](https://www.techradar.com/ai-platforms-assistants/gemini/google-launches-nano-banana-pro-a-massive-leap-in-ai-image-editing-powered-by-gemini-3-pro?utm_source=chatgpt.com)
* [The Times of India](https://timesofindia.indiatimes.com/technology/tech-news/google-launches-gemini-3-pro-image-based-nano-banana-pro-ai-model-all-details/articleshow/125469483.cms?utm_source=chatgpt.com)
* [Android Central](https://www.androidcentral.com/apps-software/ai/googles-november-gemini-drop-adds-gemini-3-nano-banana-pro-and-more?utm_source=chatgpt.com)
[1]: https://blog.google/products/gemini/gemini-3/?utm_source=chatgpt.com "A new era of intelligence with Gemini 3"
[2]: https://ai.google.dev/gemini-api/docs/models "Gemini models | Gemini API | Google AI for Developers"
[3]: https://www.reuters.com/business/media-telecom/google-launches-gemini-3-embeds-ai-model-into-search-immediately-2025-11-18/?utm_source=chatgpt.com "Google launches Gemini 3, embeds AI model into search ..."
[4]: https://ai.google.dev/gemini-api/docs/gemini-3 "Gemini 3 Developer Guide | Gemini API | Google AI for Developers"
[5]: https://deepmind.google/models/gemini/ "
Gemini 3 -
Google DeepMind
"
[6]: https://felloai.com/2025/11/the-best-ai-of-november-2025/?utm_source=chatgpt.com "The Best AI of November 2025: Gemini 3 vs GPT-5.1 ..."
[7]: https://www.techradar.com/ai-platforms-assistants/gemini/google-launches-nano-banana-pro-a-massive-leap-in-ai-image-editing-powered-by-gemini-3-pro?utm_source=chatgpt.com "Google launches Nano Banana Pro, a massive leap in AI image editing powered by Gemini 3 Pro"
[8]: https://www.androidcentral.com/apps-software/ai/googles-november-gemini-drop-adds-gemini-3-nano-banana-pro-and-more?utm_source=chatgpt.com "Everything Google added to the Gemini app in November, from Gemini 3 to Nano Banana Pro"
[9]: https://modelcards.withgoogle.com/assets/documents/gemini-2-flash.pdf?utm_source=chatgpt.com "Gemini 2.0 Flash"
[10]: https://developers.googleblog.com/new-gemini-api-updates-for-gemini-3/?utm_source=chatgpt.com "New Gemini API updates for Gemini 3"