Google Gemini: What Changed at I/O 2026 and What Matters

Google Gemini was one of the most searched terms of the week — and not by chance. Between 19 and 20 May 2026, Google I/O unleashed a sequence of announcements that change what the assistant does, where it runs, and how much it costs. This post organises what really matters, separating technical news from marketing noise.

TL;DR

• Gemini 3.5 Flash is the new main model: it surpasses Gemini 3.1 Pro in code, agentic tasks, and multimodality, at less than half the cost of frontier rivals.
• Gemini Omni generates content from any input — starting with video, with built-in SynthID watermark.
• Android has become an intelligence system: Gemini now operates natively on phones, watches, laptops, and cars.
• The new usage limits based on computation, activated on 20 May 2026, have drawn complaints from paying subscribers.
• The Google Gemini app has surpassed 900 million monthly active users, more than double a year ago.

Why Google Gemini became the talk of the week

On 19 May 2026, in Mountain View, Google opened its I/O — the annual developer conference. The central theme was one: turning every company product into an extension of artificial intelligence. Google Gemini was at the centre of virtually all announcements, and that is what triggered the search volume you probably noticed.

I have been following AI launches closely for years, and the reading here is straightforward: Google has stopped treating Google Gemini as a standalone product. It is now the layer that stitches together Android, Search, development tools, and hardware. When a company makes that move, the term becomes a search trend because millions of people want to understand, at the same time, what changed in the app they already use.

To put that interest in perspective: the Gemini app passed 900 million monthly active users, more than double a year ago, with daily requests growing sevenfold over the period. Search's AI Mode surpassed 1 billion monthly users. When a base that size receives several new models at once, curiosity turns into a search spike.

There are three fronts worth your attention, and I will go through each separating fact from promise: the new models, the new relationship between Google Gemini and Android, and a controversial change in usage limits. All data in this post comes from the official Google I/O 2026 announcements.

Gemini 3.5 Flash: the model that opens the new generation

The main announcement of the event was Gemini 3.5 Flash, the first of a new line of models that Google describes as the union of frontier intelligence with action capability. In plain English: it is a fast, cheap model that also performs autonomous tasks — browsing, writing code, chaining tools — without relying on the house's heavier model.

What the benchmarks say

According to Google, Gemini 3.5 Flash surpasses Gemini 3.1 Pro in code tests, multimodal reasoning, and agentic tasks. The released numbers were 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, and 83.6% on MCP Atlas. The model delivers output tokens about four times faster than other frontier models, at less than half the cost.

Benchmark is not user experience, and the usual scepticism applies — test numbers do not always translate into real gains for your task. But the direction is consistent: with this launch, Google Gemini pushes the fast model upward, instead of reserving high intelligence only for the Pro line. That changes the cost equation for anyone using AI at scale.

It is worth translating the term 'agentic', which recurs in the announcements. An agentic model does not just respond: it executes steps in sequence to complete a goal — open a page, read the content, fill a form, check the result. It is the difference between asking for an answer and delegating a task. Gemini 3.5 Flash was positioned exactly in that category, and that is why Google insists on tests like Terminal-Bench, which measure execution, not just knowledge.

Where Gemini 3.5 Flash is already available

The model started arriving on the day of the announcement: in the Gemini app, in Search, in the Antigravity 2.0 development environment, and in the Gemini API. For developers, it also appears in Google AI Studio and Android Studio. In other words, you can test it now, without a waiting list — and that immediate availability is part of the reason for the search spike.

Gemini Omni: when AI starts generating video

The second model novelty is Gemini Omni, a separate line that combines Gemini reasoning with content generation. The first version, Gemini Omni Flash, accepts image, audio, video, and text as input — and returns video as output.

The detail that sets it apart from previous generators is the attempt to anchor the result in the real world. Google claims the model combines factual knowledge with a sense of physics, so that the generated video respects basic rules of motion and gravity. All content comes out with the SynthID watermark — invisible to the eye but detectable by verification. It is a provenance signal that Google says has already been checked 50 million times worldwide.

This watermark point is not a technical detail. With AI-generated video becoming increasingly realistic, knowing whether a piece is synthetic becomes a matter of trust — for the press, for courts, for any company that publishes content. SynthID does not prevent misuse, but it creates a verifiable trail. It is a step in the right direction, though far from solving the problem alone.

In practice, Gemini Omni Flash appears in the Google Gemini app, as well as in Google Flow and the YouTube Shorts remix feature. For content producers, it is the most visible novelty from I/O. For businesses, it is still early: AI-generated video has real use in drafts and prototyping, but requires human review before becoming published content. Verifiable provenance is not the same as approved quality.

Android becomes an 'intelligence system'

Google repeated a phrase on stage until it stuck: Android is no longer an app platform; it has become an intelligence system. In translation, Google Gemini now operates natively in the system, not as an app you open and close. It spans phones, watches, laptops, and cars.

Googlebooks: laptops with built-in Gemini

The most concrete hardware bet is Googlebooks — a new category of premium laptops built on Android and powered by Gemini. The announced manufacturers are Acer, ASUS, Dell, HP, and Lenovo, with arrival expected in the second half of 2026. It is Google's attempt to put the assistant at the centre of the work experience, not just the phone.

Android XR glasses and Gemini Spark

Google also confirmed that Android XR glasses are arriving this year, with partners such as Samsung, Xreal, Warby Parker, and Gentle Monster. Alongside comes Gemini Spark, entering beta for Ultra plan subscribers the week after the event. The promise is an assistant that sees what you see. It is the most futuristic front of the announcement — and the one I would treat with the most caution, until hardware is in real people's hands.

This move has a clear parallel with the education sector: when intelligence becomes part of the container, not a loose resource, the user experience changes level. That was the argument we detailed when comparing the customised Moodle app with the generic one — brand and technology at the centre, not the periphery.

The new usage limits that annoyed subscribers

Not every Google Gemini novelty was well received. On 20 May 2026, Google activated a new usage limit system that replaced fixed daily prompt quotas. The reaction from some paying subscribers was negative — and it is worth understanding why.

How the new system works

The limit stopped counting prompts and started counting computation. The system weighs the complexity of each request, the activated resources — image generation, extended reasoning — and the accumulated session history. The quota renews every five hours, with a weekly cap on top.

There is a positive point: those who exceed the limit are not blocked. The system automatically downgrades the request to a smaller model and continues responding. Access does not stop — it degrades in quality.

The main complaint: opacity

The problem that concentrates criticism is the lack of predictability. In the old model, you knew how many prompts you had left. Now, 'computation used' offers no intuitive reference before you send the request. A subscriber on the AI Pro plan reported that a single prompt consumed 13% of the available quota. Others recounted that fewer than five consecutive requests — summarising a document, debugging a code snippet, generating an image — exhausted half of the five-hour quota.

For those paying between US$8 and US$250 per month, the complaint is legitimate: the limit may make sense, but it needs to be visible before sending, not after. It is the kind of change that directly affects those who use Google Gemini as a daily work tool. The updated rules are in the official Gemini help centre.

For those who depend on Google Gemini at work, there is a way to reduce the impact while the system remains less transparent: concentrate heavy requests — image generation, long document analysis — in one block, and leave simple questions for the smaller model. Distributing the load throughout the day helps avoid exhausting the five-hour quota all at once.

Comparison: Gemini 3.1 Pro, 3.5 Flash, and 3.5 Pro

The Google Gemini line now has three versions on the radar, and it is worth placing them side by side to situate where Gemini 3.5 Flash fits. The table summarises what Google released — remembering that 3.5 Pro has not yet been released to the public.

Model	Position	Status as of 20/05/2026	Highlight
Gemini 3.1 Pro	Previous generation, focus on reasoning	Available	Baseline for new benchmarks
Gemini 3.5 Flash	Fast, cheap, and agentic	In app, Search, and API	Surpasses 3.1 Pro in code and autonomous tasks
Gemini 3.5 Pro	Top of the new generation	In testing, expected June 2026	Already used internally by Google

The practical reading is simple. For most everyday uses, Gemini 3.5 Flash already suffices — and it is what you will find by default in the app. Gemini 3.5 Pro should interest those who need the maximum reasoning limit for long, complex tasks; it can only be truly evaluated when it is released, expected in June 2026.

What these novelties change in practice for businesses

Taking out the natural excitement of a launch event, what should a Brazilian manager do with all this? I separate the answer into two lists: when to take advantage now and when to hold back anxiety.

When it is worth adopting now

• Code and automation tasks: the agentic Gemini 3.5 Flash is mature enough to accelerate development and prototyping. It is worth testing via the Gemini API before committing an entire workflow.
• Team productivity: if your team already uses the Google Gemini app daily, the speed gain from the new model is immediate and requires no extra configuration.
• Draft content: Gemini Omni helps generate initial versions of video and image to quickly validate an idea before investing in production.

When to hold back anxiety

• New hardware: Googlebooks and Android XR glasses are not in anyone's hands yet. Corporate purchasing decisions wait for independent reviews.
• Critical quota-dependent workflows: if your team relies on heavy, continuous usage, test the new limit system before migrating sensitive processes to the paid plan.
• Generated video in production: without human review, do not publish. The SynthID watermark solves provenance, not quality.

A principle I apply to any technology decision is worth noting: the right tool depends on the size and risk of the operation, not on what is newest. It was the same reasoning we used when comparing the WhatsApp Business App and the Official API — adopting the most advanced is not always adopting the most suitable. And when automation comes into play, the cost model weighs as much as capability, as we discussed in the case of unlimited agents on WhatsApp.

Conclusion: what to do now with Google Gemini

I/O 2026 made it clear that Google Gemini is no longer a chatbot — it is the intelligence layer that Google wants present everywhere: on the phone, on the laptop, in search, and in code. Gemini 3.5 Flash is the novelty you can already use today and feel a difference. Gemini Omni and the new hardware are promises that deserve monitoring, not immediate decisions. And the new usage limits are the reminder that a cheap or free product always has a cost — sometimes paid in predictability.

My practical advice is straightforward: open the Google Gemini app and test Gemini 3.5 Flash on a real task from your routine, comparing it with what you used before. The best way to separate useful novelty from marketing noise is to measure it on your own use case.

If your company is evaluating how to fit AI into customer service, content, or educational platforms, it is worth talking to those who implement this type of integration daily. At Agathas Web, that is exactly the work — translating frontier launches into business decisions that pay off.