Google has unveiled its next major AI model upgrade – Gemini 3, bringing improvements in reasoning, multimodal understanding, coding, interface generation, and long-horizon agentic actions. The rollout includes Gemini 3 Pro, a new Deep Think mode, updated developer tools, and the first release of Gemini Agent for multi-step automation.
CEO Sundar Pichai said Gemini has grown into one of Google’s most ambitious scientific and product programs, benefiting from full-stack integration across research, models, infrastructure, and apps. He also shared key milestones:
- Gemini powers several major Google products.
- AI Overviews now serve nearly 2 billion users monthly.
- Gemini app has 650 million monthly users.
- Over 13 million developers use Google’s generative AI tools.
Pichai said Gemini 3 focuses on deeper reasoning, stronger intent comprehension, and better multi-step interpretation – and is launching across Search, Gemini app, AI Studio, Vertex AI, and Google Antigravity from day one.
Google claims Gemini 3 Pro delivers clearer responses, better grounding, and stronger performance across benchmarks:
- 1501 Elo on LMArena
- 37.5% on Humanity’s Last Exam (no tools)
- 91.9% on GPQA Diamond
- 81% on MMMU-Pro
- 87.6% on Video-MMMU
The model also reduces generic phrasing and improves multimodal understanding across text, images, video, audio, and code.
Gemini 3 introduces Deep Think, a mode designed for deeper step-by-step reasoning.
Benchmarks include:
- 41.0% on Humanity’s Last Exam
- 93.8% on GPQA Diamond
- 45.1% on ARC-AGI-2 (with code execution)
Google says Deep Think is undergoing extended safety reviews before full rollout.
With a 1M-token context window, Gemini 3 can now:
- Convert handwritten multilingual notes into documents
- Summarize long videos, lectures, or research papers
- Create flashcards, diagrams, or mini-tools
- Analyze sports videos
- Generate simulations, visuals, and layouts
These can be done directly inside AI Mode in Search.
Gemini 3 enhances instruction-following, zero-shot coding, and automated code execution.
Benchmark improvements:
- 1487 Elo on WebDev Arena
- 76.2% on SWE-bench Verified
- 54.2% on Terminal-Bench 2.0
Developers can access Gemini 3 via:
- AI Studio
- Vertex AI
- Gemini CLI
- Google Antigravity
Antigravity is a new development environment where Gemini can:
- Plan tasks
- Write and run code
- Execute terminal commands
- Validate outputs in a virtual computer
It integrates Gemini 3 Pro, Gemini 2.5 Computer Use, and Nano Banana (Gemini 2.5 Image).
Two new interface-generation tools are rolling out:
- Visual Layout – magazine-style visual structures
- Dynamic View – interactive UIs generated via agentic coding
Canvas also gains improved “vibe coding” for richer app generation.
Gemini Agent, built using insights from Project Mariner, can now handle tasks that require multi-step planning:
- Organize Gmail inboxes
- Prioritize or draft email replies
- Manage Google Calendar
- Extract travel details
- Compare options and generate summaries
The agent confirms sensitive actions such as payments or messages.
It also improves long-horizon planning, leading benchmarks like Vending-Bench 2.
Gemini 3 undergoes extensive internal and external evaluation with UK AISI, Apollo, Vaultis, and Dreadnode.
Enhancements include:
- Reduced sycophancy
- Better protection against prompt manipulation
- Stronger cybersecurity defenses
Availability
Gemini 3 Pro
- Rolling out globally in the Gemini app & AI Mode in Search
- Available for Google AI Plus, Pro & Ultra subscribers
- U.S. college students get 1-year Google AI Pro free
Gemini 3 Deep Think
- In extended safety review
- Coming soon for Google AI Ultra users
Developers
Access via:
- Gemini API
- AI Studio
- Vertex AI
- Gemini CLI
- Antigravity
Enterprise
- Available via Vertex AI and Gemini Enterprise
Generative Interfaces
- Visual Layout & Dynamic View rolling out gradually
- Google confirmed additional Gemini 3 series models will launch soon.
