It started with a confession. Sundar Pichai, Google’s chief executive, stepped onto the Shoreline Amphitheatre stage and told a crowd of thousands that his own engineers had been quietly depending on a new model for months, one that dramatically accelerated how Google builds its own products internally. That model was Gemini 3.5 Flash, and on Tuesday it became available to everyone.
The headline claim is striking: Gemini 3.5 Flash is not only faster than all comparable frontier models, running at four times the output token speed of rivals, it also outperforms Google’s own previous generation, Gemini 3.1 Pro, across coding, agentic reasoning, and multimodal understanding benchmarks. That is a meaningful inversion.
Historically, speed and quality have been a trade-off in AI development. With Flash 3.5, Google insists it has collapsed the gap.
Under the hood, the model supports native processing of text, images, audio, and video simultaneously while maintaining near real-time conversational responses. Google says it is designed for “long-horizon workflows” the kind of multi-step tasks that require an AI to hold a complex problem in mind and act across several tools or documents over time. Developers building on the Gemini API, as well as Google’s own Antigravity coding platform, are already running on Flash 3.5.
“When compared to 3.1 Pro, Flash is better across the board, the new model has been a game-changer for us internally at Google.”
— Sundar Pichai, Google I/O Keynote
To demonstrate this, a DeepMind engineer named Varun Mohan walked the audience through a live demo: using Antigravity and Gemini 3.5 Flash together, a team built a functioning operating system from scratch, in twelve hours. The crowd applauded. The AI community paid close attention.
