Google presented new AI models and personal agents during the I/O 2026 event. The company seeks to offer faster and more practical tools for millions of users. Highlights include the Gemini 3.5 family and features that simulate the physical world.
The main release was Gemini 3.5 Flash. Ele becomes the default model in the Gemini app and Search AI mode worldwide. Sundar Pichai, CEO of Alphabet, described the model as remarkably fast. The lighter version delivers advanced capabilities at a lower cost than equivalent competitor models.
https://twitter.com/Google/status/2056789235500466273?ref_src=twsrc%5Etfw
Gemini 3.5 Flash gains central position in services
The Gemini 3.5 Flash has a token generation speed up to four times faster than rival frontier models. Usuários notice faster responses in daily tasks. The model also strengthened defenses against harmful content. Ele mistakenly refuses fewer secure queries.
The Google positions the Flash as an ideal choice for AI agents and long workflows. Desenvolvedores gains better coding support and interactive interfaces. The Pro version of the 3.5 family is still undergoing internal testing and should reach the public in June.
- Velocidade superior in coding benchmarks and agentic tasks
- Preço about half or a third of comparable models
- Disponível globally on the Gemini app starting this Tuesday
- Ênfase on security and reducing inappropriate responses
- Integração with Search, Flow and YouTube Shorts
Omni simulates physical environments and expands video creation
The company introduced Omni, a world model designed to understand physics and predict outcomes of actions. Ele works with images, audio and video. Usuários can edit clips asking for changes in action, inclusion of characters or objects.
Omni Flash is now rolling out to Google AI Plus, Pro and Ultra subscribers. Ele replaces Veo in some tools. Aplicações include robot training and realistic content generation for Shorts. Google highlights character consistency and narrative logic in the scenes created.
The model accepts multimodal inputs and generates varied outputs. Right now, the initial focus is on video and conversational editing. Expansões for other formats will follow in the coming weeks.
Gemini Spark acts as an always-on personal agent
Gemini Spark is the new general purpose agent within the Gemini app. Ele operates in the cloud and performs tasks while the user performs other activities. The feature is in beta and is coming first to trusted testers and Ultra subscribers next week.
The agent reasons about information from connected apps and acts under the user’s direction. Google wants to make navigating digital life easier with fewer manual commands. Integrações with third-party tools will be announced soon.
Spark runs 24 hours a day in a virtual machine. Ele helps with emails, planning, and flows that require continuity. The company sees agents as a way to deepen the use of existing products.
Evento reinforces strategy against fierce competition
The Google I/O occurred amid elevated valuations for OpenAI and Anthropic. Ambas prepares possible IPOs. Pichai highlighted the company’s broad portfolio, which combines models with services for a massive user base.
Investimentos in capital grew, and the market demands deeper integration. Agentes and world models emerge in response. Omni, for example, already supports DeepMind research in robotics.
The Gemini app has more than 900 million active users, with two-fold growth in the last year. Monthly Processamento tokens reach 9.7 trillion.
Ficha technical of major releases
- Gemini 3.5 Flash: light, fast, standard model in the app and Search
- Gemini Omni Flash: world simulation, video editing, multimodal support
- Gemini Spark: personal agent in the cloud, beta for Ultra
- Disponibilidade: Flash already active; Spark next week; Pro in June
- Preços and access: varies by AI plan Plus, Pro and Ultra
Google also mentioned updates to the app redesign and new capabilities in Workspace. The focus remains on making AI more useful in everyday life without requiring extra effort from users.