I/O 2026: The Dawn of the Agent-Centric Gemini Era

At I/O 2026, Google made clear its direction to embed AI more deeply across products and infrastructure. Monthly tokens processed surpassed 3.2 quadrillion, a 7x increase year-over-year, with over 8.5 million developers using the models and the API processing 19 billion tokens per minute. Over the past 12 months, more than 375 Google Cloud customers each consumed over 1 trillion tokens.

Search and the Gemini app also showed notable growth. AI Overviews surpassed 2.5 billion monthly active users, AI Mode surpassed 1 billion, and the Gemini app grew to 900 million. Google is expanding personalization and creative demand with Personal Intelligence and the Nano Banana image generation model in the lead.

Conversational AI embedded within products is also expanding.

• Following Ask Maps, Google is testing Ask YouTube, planning wide deployment in the US this summer.
• Docs Live offers a voice-based flow for creating and editing documents, and around the same time voice features will be added to Gmail and Keep.

On the infrastructure side, Google is significantly increasing investment to sustain demand from billions of users and enterprises. This year's capex is expected to be $180 billion to $190 billion, and the company unveiled new TPU 8t and TPU 8i chips. The 8t is designed for large-scale pretraining, delivering nearly 3x the raw compute performance of the previous generation, and with JAX and Pathways, it enables training distributed across over 1 million TPUs spanning multiple sites. The 8i is designed for inference, focused on reducing latency, and both chips achieve up to 2x improvement in performance per watt.

In the model lineup, Gemini Omni and Gemini Omni Flash stand out. Omni is a world model family that generates samples in any output modality from any input, initially supporting video output before expanding to images and text. Omni Flash is available starting today in the Gemini app, Google Flow, and YouTube Shorts, with developer and enterprise APIs coming soon.

On the transparency front, Google is expanding SynthID and Content Credentials. SynthID has so far watermarked over 100 billion images and videos and 60,000 years worth of audio, and is now extending verification coverage to Search and Chrome. OpenAI, Kakao, and ElevenLabs have also adopted SynthID.

Finally, Gemini 3.5 Flash was announced. Google stated that this model outperforms 3.1 Pro on most benchmarks, with especially large gains in agentic coding, long-horizon tasks, and real-world workflows. The company emphasized that its output speed is 4x faster than other top-tier models.