Google has introduced Gemini 3, the next generation of its AI model series, bringing upgrades in long-form reasoning, multimodal interpretation, interface generation, developer tools, and agent-based task execution.
The rollout includes Gemini 3 Pro, a new Deep Think mode, expanded multimodal learning workflows, redesigned app capabilities, and the first release of Gemini Agent for multi-step automation.
Sundar Pichai Explains Gemini’s Growth and Direction
Google and Alphabet CEO Sundar Pichai shared a brief note highlighting the progress of the Gemini program, which began nearly two years ago. He said Gemini has grown into one of Google’s largest scientific and product efforts, supported by an integrated full-stack approach combining infrastructure, research, models, and products.
Key points he highlighted
- Gemini now powers multiple Google products at large scale.
- AI Overviews reach around 2 billion monthly users.
- The Gemini app has roughly 650 million monthly users.
- More than 13 million developers are building with Google’s generative AI tools.
- Each Gemini generation added improvements — multimodality, longer context, higher-level reasoning, and agentic capabilities.
Pichai said Gemini 3 brings together advancements from earlier versions to deliver deeper reasoning, better intent understanding, and more accurate multi-step interpretation with fewer instructions. He confirmed that Gemini 3 is rolling out across Google’s ecosystem — AI Mode in Search, the Gemini app, AI Studio, Vertex AI, and Google Antigravity — marking the first time a Gemini release is launching inside Search on day one.
He added that Gemini 3 represents the next phase of Google’s AI roadmap, with continued focus on intelligence, agentic systems, and personalization in future releases.
Enhanced reasoning and benchmark gains
Google says Gemini 3 Pro delivers stronger reasoning, clearer responses, and better multimodal grounding. Benchmark results include:
- 1501 Elo on LMArena
- 37.5% on Humanity’s Last Exam (no tools)
- 91.9% on GPQA Diamond
- 23.4% on MathArena Apex
- 81% on MMMU-Pro
- 87.6% on Video-MMMU
- 72.1% on SimpleQA Verified
The model is optimized to avoid generic phrasing, offer more direct answers, and maintain accuracy across text, audio, images, video, and code.
Gemini 3 Deep Think
Deep Think extends step-by-step reasoning with stronger analytical performance.
Benchmark results:
- 41.0% on Humanity’s Last Exam (no tools)
- 93.8% on GPQA Diamond
- 45.1% on ARC-AGI-2 (with code execution, ARC Prize verified)
Deep Think is undergoing extended safety reviews before wider rollout.
Learning and multimodal workflows
Gemini 3 supports expanded learning tasks through multimodal understanding and a 1M-token context window, enabling:
- Converting handwritten multilingual notes into structured documents
- Summarizing long videos, lectures, or research papers
- Creating flashcards, diagrams, or small tools
- Sports-video analysis with detailed performance breakdowns
- Generating simulations, visuals, and layouts directly inside AI Mode in Search
Coding and developer improvements
Gemini 3 improves instruction adherence, zero-shot coding, and agentic coding.
Benchmark highlights:
- 1487 Elo on WebDev Arena
- 54.2% on Terminal-Bench 2.0
- 76.2% on SWE-bench Verified
Developers can access Gemini 3 via:
- Google AI Studio
- Vertex AI
- Gemini CLI
- Google Antigravity
Google Antigravity: Agent-first development
Antigravity offers a development environment where Gemini 3 can:
- Plan tasks
- Write and execute code
- Run terminal commands
- Validate outputs using a browser-based virtual computer
It integrates:
- Gemini 3 Pro
- Gemini 2.5 Computer Use
- Nano Banana (Gemini 2.5 Image)
This enables parallel, consistent end-to-end software workflows.
Generative interfaces in the Gemini app
Two interface-generation modes are rolling out:
- Visual Layout: Creates structured, magazine-style arrangements with images and modular sections.
- Dynamic View: Uses agentic coding to generate custom interactive interfaces.
Google also notes that Gemini 3 delivers its best-ever vibe coding performance inside Canvas, enabling more feature-rich app generation within the workspace. These capabilities are part of the revamped Gemini app, which now includes a “Thinking” model selector and a My Stuff library for saved outputs.
Gemini Agent for multi-step automation
Gemini Agent, built using insights from Project Mariner, can break down and execute complex tasks. It can:
- Organize Gmail inboxes
- Prioritize emails
- Draft replies
- Manage Google Calendar
- Extract travel details from emails
- Compare options and generate summaries or plans
The agent confirms sensitive actions such as purchases or message sending.
Long-horizon planning
Gemini 3 strengthens multi-step planning and avoids tool-use drift during long tasks. It leads Vending-Bench 2, designed to test year-long operational decision-making.
Safety improvements
Gemini 3 includes stronger protections through extensive internal and external assessments.
Key enhancements:
- Reduced sycophancy
- Better resistance to prompt manipulation
- Stronger cyberattack defenses
- External reviews with the UK AISI, Apollo, Vaultis, and Dreadnode
Availability
Gemini 3 Pro
- Rolling out globally in the Gemini app and AI Mode in Search
- Available to Google AI Plus, Pro, and Ultra subscribers
- U.S. college students get 1 year of Google AI Pro free
Gemini 3 Deep Think
- In extended safety review
- Coming soon to Google AI Ultra users
Developers
- Access via Gemini API, AI Studio, Vertex AI, Gemini CLI, and Antigravity
- Supported on Cursor, GitHub, JetBrains, Manus, Replit
Enterprise
-
Available via Vertex AI and Gemini Enterprise
Generative Interfaces
- Visual Layout and Dynamic View rolling out gradually
- Some users may receive only one feature initially
Google says additional Gemini 3 series models will be released soon, and the team says it looks forward to user feedback during the rollout.