Google held its developer event "Google I/O 2026," where CEO Sundar Pichai declared in his keynote that the company has entered "the agentic Gemini era"[1]. The event featured a wide range of announcements spanning Google's product lineup, including a new language model called "Gemini 3.5 Flash," "Gemini Omni," which can produce output in any modality, and "Gemini Spark," an always-on agent that takes action on a user's behalf[1].

Entering the agentic Gemini era, with usage scaling rapidly

Google pointed to the volume of "tokens" processed — the smallest units of data its models handle — as a gauge of how widely AI is being adopted. Token processing stood at 9.7 trillion per month two years ago, grew to roughly 480 trillion at last year's I/O, and has now reached more than 3,200 trillion per month, a sevenfold increase[1].

The growth in users and developers is just as striking. More than 8.5 million developers build apps and new experiences with Google's models each month, and the model APIs now process roughly 19 billion tokens per minute[1]. Google has 13 products with more than 1 billion users each, 5 of which have more than 3 billion. The conversational Gemini app has more than doubled in a year, from 400 million monthly active users to over 900 million, and more than 50 billion images have been created with the Nano Banana image-generation model[1].

The new models "Gemini 3.5 Flash" and "Gemini Omni"

At the center of the keynote was the new model "Gemini 3.5 Flash." Positioned as the first model to combine frontier-level intelligence with the ability to take action, it shows major improvements in coding and long-horizon tasks[1]. Its output speed is four times that of other frontier models, and it is priced at less than half the cost of comparable models. Google processes more than 3 trillion tokens a day across its internal AI development tools, and it estimates that if major companies shifted 80 percent of their workloads to this model, they could save more than 1 billion USD (about 160 billion yen) per year[1]. A higher-end version, "Gemini 3.5 Pro," is due next month.

Another highlight is "Gemini Omni." This model can generate output in any modality from any input, starting with video output and expanding to images and text over time[1]. The first release, "Gemini Omni Flash," became available the same day and can be tried in the Gemini app, Google Flow, and YouTube Shorts.

The always-on agent "Gemini Spark" and Antigravity 2.0

Pichai placed particular emphasis on "Gemini Spark," a personal AI agent. Operating inside the Gemini app, it carries out tasks across a user's digital life under their direction[1]. Because it runs on dedicated virtual machines on Google Cloud around the clock, there is no need to keep a device open. Built on Gemini 3.5 and the "Antigravity" agent harness, it is set to connect with third-party tools through a mechanism called "MCP" within the coming weeks[1]. Rollout to a limited set of testers begins this week in the U.S.

For developers, Google announced "Antigravity 2.0," which has evolved beyond a coding environment into a platform for managing autonomous agents[1]. It is a new desktop application for interacting with agents, and the version of Flash optimized for it is said to run 12 times faster than other frontier models.

Agentic Search and massive infrastructure investment

Search is changing as well. Google plans to introduce "information agents" that work around the clock to find information based on a user's stated interests and help them take action, and it intends to offer — free of charge this summer — a capability that uses Gemini 3.5 Flash to generate dynamic layouts and interactive visuals for each query[1].

To support this scale, Google is sharply expanding its infrastructure investment. Capital expenditure, which was 31 billion USD (about 4.9 trillion yen) per year in 2022, is expected to reach roughly 180 to 190 billion USD (about 29 to 30 trillion yen) this year, about six times that level[1]. At the core are Google's custom chips: the 8th-generation TPU splits roles between the "8t" for training and the "8i" for inference, with the 8t offering about three times the compute of the previous generation and both chips delivering up to twice the performance per watt[1].

Efforts toward transparency are advancing too. "SynthID," the watermark applied to AI-generated content, has now been applied to more than 100 billion images and videos, and OpenAI, Kakao, and Eleven Labs were announced as new adopters[1].

※1 USD = 159 JPY (as of May 29, 2026)

Summary

Google I/O 2026 made clear the company's pivot toward agentic AI across its entire stack — from flagship products such as Search and the Gemini app to custom silicon and massive capital spending. With the new Gemini 3.5 Flash model and the always-on Spark agent at its core, how close Google can bring agents that take over users' work to practical, everyday use looks set to be the focus going forward.

Source: https://blog.google/innovation-and-ai/sundar-pichai-io-2026/