Summary of the Early Morning Google I/O 2026 Developer Conference

Just now, Google wrapped up their product launch event.

Looking back over the past six months, the AI circle has been buzzing with activity that has little to do with Google.

Claude has been on a tear, and OpenAI goes without saying.

Google has been so quiet that people almost forgot about it.

But those who know Google understand that they like to save up all their announcements and release them in one go at I/O.

This year, it finally came.

As usual, I’ve broken it down into six major sections: AI Models, Gemini Products, Visual Generation, Agent Systems, Google Search, and Other.

Without further ado, let’s dive in.

I. AI Models

• Gemini 3.5 Flash

One of the stars of this year’s I/O.

Typically, the Flash series is the lightweight, fast version, focusing on being cheap and quick, while Pro is the full flagship.

But nowadays, it’s common for the next-gen small model to outperform the previous-gen large model. So, this time around, 3.5 Flash’s capabilities in coding, agent tasks, and tool calling significantly surpass the previous 3.1 Pro.

On the Terminal-Bench 2.1 coding test, 3.5 Flash scored 76.2%, while 3.1 Pro only got 70.3%. On GDPval-AA, which measures real-world economic value tasks, 3.5 Flash achieved 1656 Elo, compared to 3.1 Pro’s 1314 Elo — a difference of over 300 points.

The benchmark scores are indeed much stronger.

However, in Humanity’s Last Exam, 3.5 Flash scored 40.2%, behind 3.1 Pro’s 44.4%, and on ARC-AGI-2, it scored 72.1%, losing to Pro’s 77.1%. These two benchmarks mainly test world knowledge and pure abstract reasoning.

This means that this time, knowledge capabilities were sacrificed in favor of enhanced task performance.

Output speed is 4x faster than other frontier models.

In terms of pricing, input is $1.50 per million tokens, and output is $9.00 per million tokens — 3x more expensive than 3 Flash but 40% cheaper than 3.1 Pro.

The trend of raising token prices across the board is really taking hold…

The knowledge cutoff is January 2025, with a context window of 1 million tokens.

As for Gemini 3.5 Pro, Sundar personally said, “Give us until next month to get it to you.” So, see you next month.

Starting today, 3.5 Flash becomes the default model for the Gemini App and AI Mode in Search, rolling out globally across API, AI Studio, Antigravity, etc. Everyone can try it out.

2. Gemini Omni Flash

This one had been buzzing on Twitter even before the conference.

To be honest, I had some expectations.

After all, Google called it “a new model that can create anything from any input.”

And right now, Google’s video model is widely considered the only one that can barely keep up with Seedance 2.0, making it the last hope for many AI animation studios.

The promotional materials looked decent.

But after trying it, I have to say it fell a bit flat.

It’s really not that great, and the Chinese accent sounds like a weird mix of Hong Kong and Taiwan accents.

I saw a comment saying: “Seems worse than Seedance 2.0.”

Not just looks — it’s also not as good as Seedance in use…

However, one feature worth mentioning is that it supports keeping a specific segment of a video unchanged while modifying the rest.

Still, considering this is the first model in the Omni family, it’s understandable that Gemini Omni Flash feels a bit lacking. Google also mentioned that Omni Pro is coming soon.

II. Gemini Products

• Gemini App Redesign

The design language of the Gemini App is officially called Neural Expressive.

On the web version, the overall color scheme has changed from the previous grayish-white interface to a blue gradient background.

At first glance, it looks quite premium, but also a bit like… phone power-saving mode?

The mobile version is similar.

The toolbar has been consolidated. Previously, uploading files, calling tools, and selecting attachments were scattered in different places; now they’re all tucked under a single “+” button.

Opening the model selector reveals a “thinking level” option at the bottom, which expands into “Standard” and “Extended.”

One thing that surprised me most was that in the settings, Google has also started implementing usage limits…

Opening the settings, I saw two progress bars: one for current usage and one for weekly limits.

They didn’t learn from Claude’s good aspects, just this one…

Currently, the new Neural Expressive design is rolling out globally on Android, iOS, and Web starting today.

• Ask Max

Google Maps is getting its biggest upgrade in a decade with a new feature called Ask Max.

You can now interact with the map using natural language.

An example from the event: a parent asked, “My child just fell into the duck pond, and the wedding starts in 30 minutes. Where can I walk to buy her a new dress?”

You could never type a question like that into the search box before, but now you can.

Google’s ecosystem is still incredibly powerful. Connecting something like Maps to Gemini can create some real synergy.

3. Ask YouTube

YouTube is rolling out a similar feature called Ask YouTube.

Instead of manually searching through videos, you can just ask it. It will provide a curated overview, tips, the most relevant video clips, or even jump directly to the most relevant part of a video.

You can also ask follow-up questions, and it remembers the context.

The concept is the same as the previous feature: turning the search box into a dialog box, whether for Maps or videos.

Ask YouTube is rolling out to Premium subscribers in the US, with a nationwide US launch this summer.

4. Docs Live

Previously, to have Gemini help you write a document, you had to craft a precise prompt and think carefully before typing.

With this feature, you don’t need to type — just speak.

It’s incredibly smooth.

The most interesting part is changing your mind mid-sentence. For example, if you say “Thursday” and then immediately correct yourself to “Friday,” Gemini automatically switches to “Friday.” Pretty neat.

It will be available to Pro and Ultra subscribers this summer. Live modes for Gmail and Google Keep will follow.

5. Gemini Live Upgrade

Voice updates for Gemini Live.

They played a few demos on stage: a Liverpool accent in English, a Haryanvi dialect in India, and a Rio de Janeiro Portuguese accent…

They switched between these three accents for a bit.

More accents will roll out over the coming weeks.

6. Daily Brief

A new feature in the Gemini App that provides a personalized summary every morning.

Starting today, it’s available to Plus, Pro, and Ultra users in the US.

7. NotebookLM

Functionality now includes cinematic video summaries. You can throw in a bunch of materials, and it will generate a narrated video with smooth animations and visuals.

Infographics have also been upgraded, with 10 preset styles to choose from: hand-drawn, cute, professional, scientific, anime, claymation…

For learning tools, flashcards and quizzes have been revamped, and progress syncs across devices.

The biggest change is that NotebookLM is now integrated with the Gemini App. The Gemini App now has a notebook feature, and notebooks created in Gemini automatically sync to NotebookLM, and vice versa.

It also supports uploading EPUB ebooks, exporting slides to PPTX format, automatically saving chat history, and generating podcasts, videos, and reports directly from conversations.

Additionally, it’s now integrated with Google Classroom, allowing college students to create their own course notebooks directly in the classroom and generate learning tools from materials provided by their teachers.

III. Agent Systems

Agents were the true main theme of this entire Google event.

• Antigravity 2.0

First, let’s talk about Antigravity 2.0.

Today, version 2.0 has finally arrived.

There are several updates.

First, a completely new standalone desktop application. Unlike before, when it was just an IDE plugin, this is now a true agent working environment.

Second, Antigravity CLI is now available globally.

This essentially replaces Gemini CLI directly.

Google’s official announcement states that after June 18, 2026, the Gemini CLI and Gemini Code Assist IDE extension will stop serving Pro/Ultra users.

Developers must migrate to Antigravity CLI.

This is important for anyone using Gemini CLI for development (though I suspect there aren’t many), so don’t wait until June 18th to find out your workflow is broken.

Third, Antigravity SDK — developers can now run the agent harness that Google uses in Antigravity directly on their own servers.

Fourth, native voice support, integrated with Gemini’s audio models, and connected to Android, Firebase, and AI Studio.

They then demonstrated on stage, using Antigravity with Gemini 3.5 Flash to build a bootable operating system from scratch.

They actually managed to create an OS that could run command lines, play Doom, and display animations.

Pretty interesting stuff.

Even more impressive is that 3.5 Flash has been specifically optimized for Antigravity. Compared to other models, it’s not 4x faster — it’s 12x faster.

Antigravity 2.0 is open globally, available to everyone starting today.

• Gemini Spark

Your personal AI agent, seemingly aimed at competing with OpenClaw.

It runs on a dedicated virtual machine in Google Cloud, 24/7. You can shut down your computer, and Spark will continue working in the cloud.

Powered by Gemini 3.5 Flash and the Antigravity harness, it can handle long-chain background tasks.

It’s also fully integrated with the Google ecosystem, helping you manage various tasks.

Spark generates real-time RSVP tracking sheets in Google Sheets, automatically syncing with Gmail. If a neighbor replies “I’m coming,” the sheet updates automatically. For neighbors who haven’t replied, it drafts reminder emails.

It also digs through your Google Drive to find the HOA rules for your neighborhood, reminding you that inflatable castles can’t be set up before Friday afternoon, and creates a party promotion deck in Google Slides…

Currently, Spark is available to some testers this week, and US Google AI Ultra subscribers can sign up for the beta starting next week.

Note: Ultra subscribers, not Pro. Honestly, who in their right mind would pay $250 a month for a Google Ultra membership? It feels like a total ripoff.

So, alongside the Spark launch, Google is completely revamping its subscription pricing system.

Previously, Google AI Ultra had only one tier at $250/month. Now it’s split into two tiers.

The new $100/month Ultra plan is aimed at developers, tech leads, and content creators, offering 5x the usage of Pro, 20TB of cloud storage, YouTube Premium, and priority access to Antigravity.

The old Ultra plan drops from $250 to $200/month, retaining all top-tier capabilities.

Spark is available on both the $100 and $200 plans.

In my opinion, Google still needs to lower its prices further.

• Android Halo

Spark works 24/7 in the cloud, but how do you see what it’s doing?

The answer is Android Halo.

Halo is a dedicated home base for agents on Android, displaying what the agent is currently doing in the status bar.

What Spark is doing, its progress, and whether it needs your confirmation, all appear in this status bar.

It’s launching later this year.

Halo was mentioned briefly, but I think it’s quite interesting and could represent a new UI layer.

Previously, Android’s UI was designed for apps, with apps being the underlying logic.

With Halo, Android is designed for agents, with agents as the underlying logic.

This could lead to many new and creative uses in the future.

IV. Visual Generation

• Google Pics

A new product in Workspace.

Pics is an image creation and editing tool for things like party flyers, infographics, and event posters.

It supports object segmentation, allowing you to select and edit any element in an image individually.

You can also edit text directly in the image and translate it into multiple languages with one click.

All outputs are automatically watermarked with SynthID to ensure traceability.

It will launch for Ultra subscribers in the US this summer.

• Stitch

Stitch is Google’s UI design tool. Over the past year, users worldwide have generated over 100 million UI screens with Stitch, and Google says they use it internally too.

(PS: Anyone here used it?)

This update includes real-time voice collaboration (speak, and the UI changes in real-time), exporting code, publishing directly to Netlify, and integration with Antigravity.

The coolest feature is Flow Music.

It’s decent, but still falls short of Suno.

So, the release logic for Flow is pretty clear.

They want to be the entry point for the entire creative workflow.

From canvas, to script, to shots, to editing, to color grading, to music — everything in one place.

But honestly, while it’s feature-rich, it’s also not very user-friendly…

4. SynthID

A small update.

Google’s AI watermarking technology, specifically designed to mark content generated by AI.

It has already watermarked over 100 billion images and videos, plus audio totaling over 60,000 years in length.

The new change is that now in Chrome, you can right-click on an image or use Circle to Search to check if an image is AI-generated.

What surprised me most was that Google announced OpenAI, Kakao, and ElevenLabs have also joined SynthID.

OpenAI also made an announcement.

This is the most story-driven detail of the entire conference.

For the past three years, these two companies have been at each other’s throats. Today, they set aside their differences to collaborate on SynthID.

The problem of AI-generated fake images, fake voices, and fake videos has become so severe that everyone has to put aside their pride and work together.

Nvidia joined last year, and Sony Pictures, Reuters, and TikTok are on their way.

V. Google Search

AI Mode now has over 1 billion monthly active users, and query volume has doubled every quarter since launch.

Today, it was officially announced that the underlying model has been upgraded to Gemini 3.5.

There are four specific updates:

1. Redesigned search bar. Google says this is the biggest upgrade to the search bar in 25 years. Previously, you could only type. Now, you can drop in images, files, and videos, and the search will understand cross-modally. It also uses AI to help complete your query, teasing out what you really want to ask.

2. AI Overviews and AI Mode have merged. The search results page now transitions naturally into conversational follow-ups, with context following you.

3. Search Agents. You can now create agents within Search. You can launch multiple agents simultaneously, letting them monitor things in the background 24/7.

4. Agentic Coding enters Search. Search can now build custom interactive interfaces from scratch in real-time based on your query. This is powered by Antigravity in the background. When searching, it calls a containerized agent environment, allowing 3.5 Flash to write and run code in real-time, embedding the rendered results back into the search results. This will be available for free to all users this summer. Embedding generative UI directly into search might be the biggest evolution of the search product since 1998.

Due to word limits, the other two updates are posted in the comments…

Final Thoughts

Finally… I’m done summarizing…

Google’s press conferences are always incredibly information-dense to the point of being overwhelming.

At the very end, Hassabis closed with a line that really moved me.

He said:

When we look back at this time, I think we’ll realize that we were standing in the foothills of the singularity.

I truly believe that.

AI, at least for now, is an amplifier of human intelligence.

Perhaps we are entering a new golden age of scientific discovery and progress.

I also hope that in the future.

We can continue to witness it, together.