Robots Atlas>ROBOTS ATLAS
Gemini 3
AI Models›Gemini

Gemini 3

3 · Family: Gemini
Gemini 3 is the third generation of Gemini models from Google DeepMind, a multimodal model family with advanced reasoning, coding, and agentic planning capabilities.
⏳ Preview⏳ Limited accessLLMMultimodalReasoning modelTool-using model📁 Gemini
Context window
1M tokenów
tokens
Parameters
nieujawnione
parameters
Max output
65,536
tokens
Release date
18 November 2025
🔬Google DeepMindResearch lab🏢GoogleOwner
Access:APIHostedDeployment:☁ Cloud

Overview

Gemini 3 is the third generation of Gemini models developed by Google DeepMind, announced on November 18, 2025. The family is designed for advanced multimodal reasoning, agentic coding, and long-horizon planning. The first model released in this generation was Gemini 3 Pro in preview.

The model natively supports multiple input modalities (text, image, audio, video, PDF documents) with a context window of up to 1 million tokens. It supports tools including function calling, code execution, and search as a tool. Access to Gemini 3 family models is available via the Gemini API, Google AI Studio, Vertex AI, and the Gemini app.

Classification
LLMMultimodalReasoning modelTool-using model
Family: Gemini
Applications
CodingContent generationDocument analysisWorkflow automationWriting assistanceQ&A / Question answeringResearch assistanceData analysis
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
📏 Context: 1M tokenów
🧩 Parameters: nieujawnione
✓ Tools
📥 Input: text, image, audio, video…
Platforms
Vertex AI

Technical specification

Context window
1M tokenów
tokens
Parameters
nieujawnione
parameters
Max output tokens
65,536
tokens per response
Knowledge cutoff
1 Jan 2025
Knowledge boundary
License
proprietary
Hardware requirements
Access via Google Cloud infrastructure (Vertex AI / Gemini API) or Gemini app.
Features:✓ Tool use
Modalities
⬇ Input
textimageaudiovideodocuments
⬆ Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
Category: reasoning
Multi-step reasoning
Category: reasoning
Long context
Category: reasoning
Coding
Category: coding
Function Calling
Category: planning
Structured output
Category: structured_generation
Audio understanding
Category: audio
Image understanding
Category: vision
Video Understanding
Category: video
Chart understanding
Category: vision
Diagram reasoning
Category: reasoning
OCR
Category: vision
Multilingual
Category: language
Planning
Category: planning
Streaming output
Category: reasoning
Interleaved Multimodal Input
Category: reasoning
Multimodal understanding
Category: multimodal
Application domains
CodingContent generationDocument analysisWorkflow automationWriting assistanceQ&A / Question answeringResearch assistanceData analysisSearch assistanceTranslationTutoring / Education

Benchmark results

14 benchmarks
Humanity's Last Exam
accuracy · No tools, Gemini 3 Pro Thinking (High)
37.5%%
📄 https://deepmind.google/models/gemini/flash/
Score for Gemini 3 Pro Thinking (High) variant, no tools. Full set (text + MM).
GPQA Diamond
accuracy · No tools, Gemini 3 Pro Thinking (High)
91.9%%
📄 https://deepmind.google/models/gemini/flash/
Score for Gemini 3 Pro Thinking (High) variant, no tools.
SWE-Bench Verified
accuracy · Single attempt, Gemini 3 Pro Thinking (High)
76.2%%
📄 https://deepmind.google/models/gemini/flash/
Agentic coding benchmark, single attempt.
MMMU-Pro
accuracy · No tools, Gemini 3 Pro Thinking (High)
81.0%%
📄 https://deepmind.google/models/gemini/flash/
Multimodal understanding and reasoning benchmark.
MMMLU
accuracy · Gemini 3 Pro Thinking (High) / Gemini 3 Flash Thinking
91.8%%
📄 https://deepmind.google/models/gemini/flash/
Multilingual Q&A benchmark. Flash: 91.8%, Pro: 91.8%.
ARC-AGI-2
accuracy · ARC Prize Verified, Gemini 3 Pro Thinking (High)
31.1%%
📄 https://deepmind.google/models/gemini/flash/
Abstract reasoning puzzles, ARC Prize Verified.
Humanity's Last Exam
accuracy · No tools, Gemini 3.1 Pro Thinking (High)
44.4%%
📄 https://deepmind.google/models/gemini/pro/
Score for Gemini 3.1 Pro Thinking (High) variant, no tools. Full set (text + MM).
GPQA Diamond
accuracy · No tools, Gemini 3.1 Pro Thinking (High)
94.3%%
📄 https://deepmind.google/models/gemini/pro/
Score for Gemini 3.1 Pro Thinking (High) variant.
SWE-Bench Verified
accuracy · Single attempt, Gemini 3.1 Pro Thinking (High)
80.6%%
📄 https://deepmind.google/models/gemini/pro/
Agentic coding benchmark, single attempt.
ARC-AGI-2
accuracy · ARC Prize Verified, Gemini 3.1 Pro Thinking (High)
77.1%%
📄 https://deepmind.google/models/gemini/pro/
Abstract reasoning puzzles, ARC Prize Verified.
ARC-AGI-2
accuracy · ARC Prize Verified, Gemini 3.1 Deep Think Feb 2026
84.6%%
📄 https://deepmind.google/models/gemini/deep-think/
Abstract reasoning puzzles, ARC Prize Verified. Best result in Gemini 3 family.
Humanity's Last Exam
accuracy · No tools, Gemini 3.1 Deep Think Feb 2026
48.4%%
📄 https://deepmind.google/models/gemini/deep-think/
Full set (text + MM).
GPQA Diamond
accuracy · No tools, Gemini 3.1 Flash-Lite High
86.9%%
📄 https://deepmind.google/models/gemini/flash-lite/
Scientific knowledge benchmark.
MMMU-Pro
accuracy · No tools, Gemini 3.1 Flash-Lite High
76.8%%
📄 https://deepmind.google/models/gemini/flash-lite/
Multimodal understanding and reasoning benchmark.

Pricing

Technical architecture

Deployment and security

☁ Available on platforms
☁Vertex AIPlatform
🔒 Security / Enterprise
✓ Verified enterprise information

Gemini 3 przeszedł kompleksowe ewaluacje bezpieczeństwa zgodnie z Frontier Safety Framework Google DeepMind. Model card dostępny publicznie pod adresem deepmind.google/models/model-cards/gemini-3-pro.

Updated: 1 May 2026↗ Security documentation

Sources and related pages