Google launched Gemini 3.5 Flash at I/O 2026, calling it the company’s strongest model for coding and autonomous AI agents. The model is four times faster than rival frontier models, can run autonomously for hours, and was co-developed with Antigravity, Google’s new agent-first IDE. Flash is now the default model across the Gemini app, AI Mode in Search, and the new Gemini Spark personal agent.TL;DR
Somewhere inside Google’s campus, an AI agent recently built an operating system from scratch. Not a toy demo, not a sanitised benchmark exercise, but a full OS assembled by software agents spawning off to handle separate components before stitching everything together. That, at least, is what Google showed the world at I/O 2026 on Monday, and it wants the demonstration to mark a turning point: the moment its AI stopped being a chatbot and started being a builder.
The model behind the feat is Gemini 3.5 Flash, which Google is calling its strongest model yet for coding and autonomous AI agents. According to Koray Kavukcuoglu, DeepMind’s chief technologist, Flash “outperforms our latest frontier model, 3.1 Pro, on nearly all benchmarks,” including coding, agentic tasks, and multimodal reasoning. The company says the model is four times faster than competing frontier models, and that an optimised version pushes that figure to 12 times faster with equivalent quality.
Speed is the point. When agentic AI systems run in production, they do not take turns. Multiple agents operate simultaneously on long-running tasks, and latency becomes a bottleneck that compounds with every parallel process. Google has built Flash specifically for that environment, co-developing it alongside Antigravity, the company’s agentic development platform and IDE.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
On stage at I/O, Google engineer Varun Mohan demonstrated the OS-building exercise inside Antigravity. Agents spawned to work on separate components, each handling a distinct piece of the system, before converging to assemble the whole. It was the kind of demo designed to make developers rethink what “coding assistant” means, and Google leaned into the implication: this is not autocomplete with ambition. It is autonomous software engineering.
Google released Antigravity 2.0 alongside the model. The update turns the platform into a standalone desktop application built around what the company calls “agent-first development,” a native environment where AI agents can, in Google’s phrasing, “live, work, and execute.” The tool includes a CLI for terminal-first developers, an SDK for custom agent behaviours, and integrations with Google AI Studio, Firebase, and Android.
Flash can run autonomously for multiple hours, according to Google, pausing to request user input only when it reaches a decision point or permission boundary that requires human judgment. That design sits in deliberate tension: the model is built for independence, but it is also built to know when to stop and ask.
Tulsee Doshi, Google’s senior director and head of product for the Gemini model, outlined the broader architecture. A forthcoming 3.5 Pro model will serve as the orchestrator and planner, while Flash handles the sub-agent work. The division is instructive. Google is not building a single all-powerful model; it is building a hierarchy of specialists, with Pro thinking and Flash doing.
Partners are already testing the approach. Banks and fintechs are reportedly automating multi-week workflows, and data science teams are using Flash agents to surface insights that previously required manual investigation. Google has not disclosed specific client names or quantified results, so those claims remain difficult to independently verify.
The launch signals a broader strategic pivot. Google is no longer positioning its AI primarily as a conversational interface. Flash is the default model in the Gemini app and powers AI Mode in Search globally. It also underpins Gemini Spark, the new personal AI agent that runs around the clock on Google Cloud infrastructure, handling tasks across Gmail, Docs, and more than 30 third-party integrations.
The timing matters. Vibe coding and AI-assisted development have moved from novelty to production tooling in the span of a year, and every major lab is racing to claim the agentic layer. OpenAI, Anthropic, and a growing roster of startups are building competing agent frameworks. Google’s bet is that raw speed, tight integration with its own cloud infrastructure, and a purpose-built IDE will give Flash an edge that benchmarks alone cannot capture.
The release also arrives under a shadow. Earlier this year, a lawsuit alleged that a man nearly carried out a mass casualty event after extended interactions with Google’s Gemini chatbot. Google says it has strengthened cyber and CBRN safeguards for the 3.5 generation and improved calibration for sensitive questions. Whether those measures are sufficient is a question regulators and courts, not just engineers, will ultimately answer.
Gemini 3.5 Flash is available now through Antigravity, the Gemini API, Gemini Enterprise, the Gemini app, and AI Mode in Search. The agentic era Google has been promising since Cloud Next has, it seems, arrived in earnest. The question is no longer whether AI can write code. It is whether anyone, human or otherwise, can keep up with how fast it is learning to build.
Get the most important tech news in your inbox each week.
