Google launches Gemini 3.5 Flash for AI agents

Google has introduced Gemini 3.5 Flash, a model designed to let AI do more than respond to messages and execute complete tasks instead. The announcement comes with new tools for creating, coordinating and deploying AI agents from the desktop, an API, mobile devices and Android.

The company says Gemini 3.5 Flash outperforms Gemini 3.1 Pro in nearly all of its benchmark tests and runs four times faster than other frontier models. Google positions it as the engine for workflows where AI must reason, use tools, write code and maintain task state across multiple interactions.

Antigravity becomes the hub for agents

The main new feature is Google Antigravity, Google's platform for creating production-ready applications from an idea. It now supports several ways to work with agents, including agents that can divide tasks among themselves and execute them in parallel.

• Antigravity 2.0: a standalone desktop application from which you can coordinate multiple agents, schedule automated tasks and connect them with Google AI Studio, Android and Firebase.
• Antigravity CLI: a terminal-based version for people who prefer to work without a graphical interface. Google recommends migrating from Gemini CLI.
• Antigravity SDK: lets you define custom behaviors for agents and host them on the infrastructure each developer chooses.
• Integration with Gemini Enterprise Agent Platform: connects Antigravity directly to Google Cloud projects for enterprise workloads.

Google has also announced a new Google AI Ultra plan costing 100 dollars per month. It includes a usage limit in Antigravity five times higher than the Google AI Pro plan. For a limited time, new and existing subscribers can claim 100 dollars in additional Antigravity credits if they reach their plan's quota. The offer ends on May 25, 2026, and is subject to conditions.

Agents managed from a single API call

Developers will be able to create managed agents through the Gemini API. With a single call, an agent can reason, use tools and execute code in an isolated Linux environment.

These agents are built on the same Antigravity infrastructure Google uses for its own products. Each interaction creates a persistent environment: in subsequent calls, the agent retains the files and session state. In practice, this lets you start a task, stop it and resume it later without rebuilding the entire context.

It will also be possible to customize agents with your own instructions and skills written in Markdown files. Google will offer templates to get started from the Google AI Studio playground.

From an idea on your phone to an Android app

Google AI Studio is also expanding its access channels. The new mobile app, available for preregistration this week, will let you capture an idea from your phone and generate a prototype to continue working on later from your computer.

The platform also adds several integrations:

• Agents can use Google Workspace APIs directly within applications.
• Projects can be exported to Antigravity with one click, preserving the context as you move from prototyping to local development and then to production.
• Native Android support lets you create applications from a prompt.
• Integration with Google Play Console lets you publish applications directly to the testing channel.

For you, the important change is that the process no longer has to start with an empty project and hours of configuration. You can describe an application, get a prototype, continue developing it on your computer and prepare a test version for Android from the same ecosystem.

A 2 million dollar prize

Google has also launched the Build with Gemini XPRIZE Hackathon, a global competition with a 2 million dollar prize pool, the largest the company has announced so far for a hackathon.

Participants will be able to create applications with Gemini models, tools and platforms to address problems such as food waste and scientific research. Finalists will present their projects in September during the Moonshot Gathering in Los Angeles.

Google's move points to a specific transition: moving from asking AI for something and waiting for an answer to delegating a multistep process to it. The question now is whether these agents can remain reliable when working with real data, tools and code, because speed is only useful when they can also complete the task without losing control.