Google DeepMind Releases Gemini 3 with Million-Token Context Window and Multimodal Reasoning
Google DeepMind's Gemini 3 introduces a 1-million token context window, native video understanding, and breakthrough multimodal reasoning capabilities.
In Brief
Google DeepMind's Gemini 3 features a 1-million token context window (50x GPT-5.6), native video understanding, and 92.4% MMLU score. Available via Google Cloud Vertex AI and Gemini Advanced at €29.99/month.
Google DeepMind has released Gemini 3, its most advanced AI model, featuring a 1-million token context window, native video understanding, and breakthrough multimodal reasoning that processes text, images, audio, and video simultaneously.
The 1-million token context window allows Gemini 3 to process entire codebases, lengthy legal documents, and full-length movies in a single inference pass. This is 50x larger than GPT-5.6's 20,000-token context and 10x larger than Claude Fable 5's 100,000-token context.
Information
Gemini 3 scored 92.4% on the MMLU benchmark — the highest ever recorded for an AI model — surpassing GPT-5.6's 89.7% and Claude Fable 5's 90.1%.
The model's native video understanding is a first for major AI models. Gemini 3 can analyze video content frame by frame, understanding temporal relationships, action sequences, and scene transitions. Google demonstrated the model reviewing a 30-minute tutorial and providing a step-by-step summary.
“Gemini 3 represents a fundamental breakthrough in AI reasoning. The million-token context window isn't just bigger — it enables entirely new categories of applications that weren't possible before.”
— Demis Hassabis, CEO of Google DeepMind
Gemini 3 is available through Google Cloud Vertex AI, Gemini Advanced (€29.99/month), and as an API with pricing at $15 per million input tokens. Google has also released a Gemini 3 Nano variant optimized for on-device deployment.
Key Takeaways
- 011-million token context window — 50x larger than GPT-5.6 and 10x larger than Claude Fable 5
- 02Native video understanding with frame-by-frame temporal analysis
- 0392.4% MMLU benchmark score — highest ever recorded for any AI model
- 04Simultaneous multimodal processing of text, images, audio, and video
- 05Available at $15/million input tokens via API and as on-device Nano variant
Frequently Asked Questions
How does Gemini 3 compare to GPT-5.6?
Gemini 3's 1-million token context is 50x larger than GPT-5.6, and it scores 92.4% on MMLU vs GPT-5.6's 89.7%. Gemini 3 also offers native video understanding that GPT-5.6 lacks.
What is Gemini 3 Nano?
Gemini 3 Nano is a compact variant optimized for on-device deployment on smartphones and edge devices, supporting up to 128K tokens of context.
Sources
Topics
Stay Ahead with TECHRADAR360
Get the latest tech news delivered to your inbox every morning.
