There are technological conferences that leave headlines about new products and there are conferences that leave headlines about a change of era. The Google I/O on May 19-20, 2026 clearly belongs to the second category. Not because Google has presented a list of impressive gadgets - that too - but because the central message of Sundar Pichai and his team drew quite precisely where the entire industry is heading: towards artificial intelligence agents that not only answer questions but execute complex tasks autonomously.
The difference seems subtle but it is not. An AI that answers questions is a sophisticated tool, just like the web search engine was at the time. An AI that acts on your behalf—that can reserve a table, fill out a form, write code, manage your inbox, coordinate multiple tools—is something qualitatively different. He is a collaborator, not a consultant.
The numbers that Pichai wanted everyone to hear
The keynote opened with a volley of figures. AI Mode in Search surpassed 1 billion monthly active users, with queries doubling every quarter. The Gemini application has exceeded 900 million monthly users. More than 8.5 million developers are building apps with Google models every month.
They are numbers designed to convey one thing: Google has not lost the artificial intelligence race. The fear that ran through the company in 2023, when ChatGPT threatened to make the search engine irrelevant, seems to have dissipated. The company has not only survived the challenge, but has turned its search—the product that has generated most of its revenue for two decades—into a conversational AI platform that remains the most used in the world.
Google came to I/O with a diagnosis about its own industry: companies adopted artificial intelligence, discovered it was expensive, and now look at the bill more closely than the product.
Gemini 3.5 Flash: the bet on cheap and fast
The model that was the protagonist of the event was not the most intelligent in the Google catalog, but rather the most efficient. Gemini 3.5 Flash was presented as the direct answer to the concern dominating the enterprise sector right now: the costs of AI are too high. Pichai said it directly on stage: there are testimonies from technology directors who have already exhausted their annual token budget and it is May.
That Google's CEO said it out loud, on the main stage of his most important conference of the year, rather than whispering it in private conversations, is significant. It means that the industry has moved from the phase of uncritical enthusiasm to the phase of financial accountability. It is no longer enough to say that AI is transformative; It must be demonstrated that the transformation justifies the monthly calculation bill.
Google presented internal data as a witness case: in March they processed half a billion tokens daily in internal use, and at the time of I/O that figure had already exceeded three billion. They are using their own cheap models at a scale that is doubling every few weeks. The implicit argument is powerful: if even Google uses low-cost versions of its own AI for most of its internal operations, perhaps companies that insist on always using the more powerful model are making financially irrational decisions.
Antigravity 2.0 and the agent platform
Beyond the models, the most relevant technical novelty of the I/O was the presentation of Antigravity 2.0, the evolution of Google's AI agent development platform. The tool allows multiple subagents to be launched in parallel to address complex workflows, with integrated sandboxing, credential masking, and enhanced code control policies.
What makes this interesting is not the technology itself — Microsoft, Anthropic and other players are developing similar capabilities — but the entire ecosystem that Google is building around it. With the new Antigravity SDK, developers have programmatic control over the agent system and can deploy it on their own infrastructure. This means that companies with strict data sovereignty or security requirements can use Google agents without their sensitive information having to pass through the Californian company's servers.
Main news from Google I/O 2026
- Gemini 3.5 Flash: fast and economical model as a new business standard
- Antigravity 2.0: agent platform with parallel subagents
- WebMCP: Open web standard for AI agents in the browser (with Microsoft, W3C)
- Android 17: Stable CLI for AI-assisted development, semantic code analysis
- Android XR glasses: compatible with iPhone and Android, real-time translation with Gemini
- Search AI Mode: surpassed 1 billion monthly active users
WebMCP: the standard that no one expected
One of the most discreet technical developments of I/O but potentially most impactful in the long term was the announcement of WebMCP, an open web standard proposed by Google together with Microsoft and incubated in the W3C Machine Learning group. The standard allows developers to expose JavaScript functions and HTML forms as structured tools that browser AI agents can execute directly.
In practical terms, it means that any website could become a tool that AI agents use autonomously, without the agent having to "simulate" the behavior of a human user by clicking on buttons. It's the difference between an agent that scrapes a website and an agent that uses a well-designed API. Faster, more reliable, less prone to errors when the website design changes.
Glasses: the hardware that Google had been trying to sell for years
The I/O also included the launch of new smart glasses under Android XR. Compatible with both Android and iPhone—in itself a notable shift in Google's strategy—they include real-time translation, navigation, calling, notification summaries, and contextual visual assistance, all powered by Gemini. To activate the assistant, simply touch the side of the mount or say "Hey Google."
XREAL also announced that it will launch its own Android XR glasses before the end of 2026. The smart glasses market begins to fill with options precisely when Meta has made its Ray-Ban with AI the first AI wearable that has had massive commercial success. Google is late to that party, but it arrives with more processing power and a considerably more sophisticated language model than its competitors.
The big absentee from I/O: Google presented everything it could present in terms of generative AI, agents and hardware. What he didn't present was a clear answer to the question of how he is going to monetize all of this in a way that justifies the billions he is investing in computing infrastructure. Gemini Flash efficiency is the tactical answer; the strategic response remains hazy.
The age of agents: promise and risk
The core concept of I/O 2026—the transition from responsive AI to acting AI—is genuinely transformative, but it also brings with it risks that the event tiptoed around quite a bit. When an AI agent has the ability to perform real actions on your behalf—sending emails, making purchases, modifying documents, interacting with external services—the question of who is responsible when something goes wrong becomes urgent and complex.
Google has built sandboxing and credential control policies into Antigravity, which is a step in the right direction. But regulation in this area is far behind technology. In Europe, the AI Act is just beginning to be implemented for high-risk systems, and autonomous agents that perform tasks in the real world fall into a regulatory gray area that lawmakers have not yet fully defined.
I/O 2026 was, in short, the conference in which Google said clearly: we are no longer in the phase of experimenting with AI, we are in the phase of deploying it on an industrial scale. If that transition goes well, Google will have preserved its dominant position in the Internet services layer. If it goes wrong, it will have been the company that built the infrastructure on which everyone will compete.
Share this article
