Gemini 3.5 is here! Tonight, Google has officially sidelined itself.

Author: New Intelligence Source
Google I/O 2026 is firing on all cylinders!
Just now, Chop and Demis Hassabis took the stage together, unveiling all the big moves they've been sitting on for half a year.
No surprises here, the star of the night is officially the Gemini Omni!
As a true 'all-rounder' model, Omni can take any form of input and generate any content. It even supports video output at launch, making it the 'video version of Nano Banana.'
Another highlight tonight belongs to Gemini 3.5 Flash.
In nearly every benchmark, the 3.5 Flash absolutely crushed its predecessor, the flagship Gemini 3.1 Pro. The output speed has also doubled, and compared to GPT-5.5 and Opus 4.7, it's over 4 times faster. The more powerful 3.5 Pro will drop next month.
Additionally, a wave of heavy new products was showcased:
Antigravity 2.0: a new independent desktop application evolved from an IDE into an agent development platform.
Gemini Spark: a personal AI agent running 24/7 in the cloud.
Gemini App revamp: codenamed Neural Expressive, now billing based on compute power.
AI Ultra subscription plan: a new $100 version, the top tier reduced from $250 to $200.
The biggest upgrade in Google Search in 25 years: integration of 3.5 Flash, with new smart search box, auto-generated mini apps, etc......
It’s no exaggeration to say that the density of valuable insights from this I/O is the highest in years.
Gemini Omni debut: an 'all-purpose' AI is born.
As the pre-release video hinted wildly, the highly anticipated Gemini Omni has finally arrived. Hassabis took the stage personally to announce, "We are taking the next important step—Gemini Omni, a new model that creates content from any input."
This lineup says it all. What Google is aiming to create this time is an 'all-purpose' AI creation engine. It merges the intelligence of Gemini with the strongest generative AI, maxing out understanding of the world, multimodal, and editing across three dimensions. To put it simply, give it any combination of images, audio, video, and text, and it can generate a high-quality video. Plus, you can edit the video in a conversational manner.
More critically, Omni isn’t just 'looking like' it understands; it genuinely grasps the physical world. Hassabis stated that previous systems often struggled when simulating concepts like gravity and momentum, but Omni achieved a 'leap change'. It infused Gemini's 'world knowledge' and 'reasoning capabilities' into video generation.
Give it a prompt 'explain protein folding with clay animation,' and the generated video accurately depicts each step of the amino acid chain folding into α helices and β folds, visually rendered as exquisite stop-motion animation.
For example, assigning corresponding objects to the 26 letters of the English alphabet. C is for capybara, D is for disco ball, L is for lava lamp. Omni doesn't just collage materials; it truly connects language, images, and semantics.
I have to say, the leap from realism to meaningfulness is monumental.
On stage, Hassabis pulled out a selfie video and started live modifying it. A circle drawn on his palm became a black hole, and a street scene from an evening stroll transformed into a cyberpunk setting. A single sentence rewrote the scene; a single sentence changed the world. Anything can become a canvas for creating a new reality. For example, playing with fire in a selfie, drawing a circle on a piece of paper immediately turns it into a black hole; all sorts of imaginative play can be realized.
Moreover, this isn’t a one-time generation. You can continue to chat. The video output from Gemini Omni maintains character consistency, logical physics, and coherent scene memory.
Starting from a raw performance scene. In the second round, 'transport the violinist into the environment of this image,' with a reference image of snowy mountains and grasslands; the scene instantly switched, with actions and light adapting perfectly to the new environment.
In the third round, 'switch the camera to behind the violinist's shoulder,' changing the perspective, but the performance actions and music remained completely continuous.
No matter how the scene changes, the main subject of the image remains intact.
What's even more mind-blowing is the flexibility of Omni's input. Images, text, video, audio—any reference material can be mixed in, generating a coherent output. You can even create your own Avatar, letting an AI version of you pop up in any scenario, speaking your voice and doing things you’ve never done.
Currently, Omni Flash has officially launched, with the API version set to open in the coming weeks. An even stronger Omni Pro is also on its way. Thanks to Google’s robust integration capabilities, Omni's debut has already connected with Gemini App, Google Flow, and YouTube Shorts, with YouTube Shorts users even able to use it for free.
Flash crushes Pro: 3.5 redefined 'flagship'.
After Gemini Omni, the highlight of the I/O conference was the launch of the brand new flagship Gemini 3.5 Flash. Google defines it as the most powerful coding and agent model to date.
On site, Hassabis announced, "3.5 Flash outperformed Gemini 3.1 Pro in virtually all benchmark tests!" Mind you, 3.1 Pro was Google’s flagship model released just three months ago, and now a Flash-level model has crushed it.
Unexpectedly, Google has delivered impressive results in such a short time:
Terminal-Bench 2.1 (coding): 76.2%
GDPval-AA (real-world agent tasks): 1656 Elo
MCP Atlas (large-scale tool usage): 83.6%
CharXiv Reasoning (multimodal understanding): 84.2%
The parameters were too abstract, so let's look at real extreme demos. In an instant, 3.5 Flash can digest a dense academic paper and write a visually stunning interactive website. In agent tasks, through Antigravity, it can complete multi-step workflows, automatically classifying and naming assets on screen. Alternatively, using two agents, it replicated the AlphaZero paper and wrote a fully functional game in just six hours.
93 agents built an OS in just 12 hours.
You can see that all the capabilities of 3.5 Flash are achieved through the new Antigravity 2.0. Today, Google’s agent development platform, Antigravity, has upgraded to 2.0, evolving from an IDE into an independent desktop application, fully embracing agent-first design.
Varun took the stage to deliver a breath-taking demo. He used Antigravity powered by 3.5 Flash to build an operating system from scratch. 93 sub-agents worked in parallel, issuing over 15,000 model requests, processing 2.6 billion tokens. After 12 hours, a completely blank project transformed into a fully functional OS kernel. Scheduler, memory management, file system—every line of code was written, tested, and audited by agents. The API cost was under $1000.
Next, he tried to run DOOM on this AI-written operating system. The first attempt failed due to missing video and keyboard drivers. So, he entered the fix commands in Antigravity 2.0 on the spot, and the agent started auto-generating driver code. After a while, the DOOM graphics appeared on the screen, and the crowd erupted.
To sum up, Antigravity 2.0 brings core upgrades including:
Sub-agents can be generated dynamically, with the main agent breaking tasks into sub-tasks to assign out, running in parallel without interference.
Asynchronous task management allows long-running operations to no longer block the main thread.
Scheduled Tasks can set 'timed tasks' for agents to execute automatically, like checking PR status daily or running a health check script hourly.
New slash commands: /goal allows the agent to run everything at once, /grill-me lets the agent clarify requirements before acting, and /browser gives explicit control over browser usage.
However, all these are capabilities already running internally. The speed at which Google processes tokens with Antigravity was 500 billion per day in March. Now, it has skyrocketed to 3 trillion per day. And this 12x accelerated version of Flash is available in Antigravity starting today.
3.5 Flash simultaneously becomes the default model for both Gemini App and Google Search AI Mode, available to users globally. Developers can access it through Antigravity 2.0, Gemini API, and Google AI Studio. Even more explosively, 3.5 Pro is currently in internal testing and will be released next month.
24/7 personal assistant: Google Spark has finally arrived.
Tonight's third major release belongs to Gemini Spark! Its positioning is very clear: your personal AI agent. Even when you close your laptop, it keeps running. It's hosted on a dedicated VM in the cloud, ensuring 24/7 availability.
Gemini Spark is powered by Gemini 3.5 + Antigravity framework, deeply integrating Google’s 'office suite'. Product VP Josh Woodward demonstrated two scenarios that sent the crowd into a frenzy.
The first scenario was a work scene: entering a command, 'Help me draft an email to the team summarizing all information about Gemini Live's release from the past week.' Spark automatically pulled information across Gmail, Docs, and chat logs, even calling upon a 'ghostwriter' skill written by Woodward himself to match the email to his personal tone. The entire process happened in the background; humans only needed to review and send. Yes, Spark supports custom skills, allowing it to learn your tone, preferences, and working style.
The second scenario was planning a block party. Once Spark received the task, it executed step by step. It created an RSVP tracking sheet in Google Sheets, directly connected to Gmail, auto-updating who responded. For neighbors who hadn’t signed up, Spark drafted reminder emails, generating drafts for confirmation before sending. Then, it even created a promotional deck in Google Slides, including info about an inflatable castle for the block. Not a single app was opened throughout the process.
Not only that, but Spark also boasts powerful voice input capabilities. On stage, Woodward pulled out his phone and directly voiced three tasks: 'Find all meetings with Sundar marked in bright pink', 'Draft an invitation for new neighbor John to join the block party list', 'Create a document listing things to do for the kids before the school year ends, sorted by deadline'.
Voice was directly converted into text commands, and Spark automatically split a continuous voice input into three independent task threads, executing them in parallel in the background.
In terms of pricing, the AI Ultra subscription is available at $100 per month for access to Spark Beta. The highest-tier Ultra plan has been reduced from $250 to $200. Spark will first open Beta testing for AI Ultra users in the U.S. next week.
That night, Google tore open the ASI gateway.
Looking back at this I/O, what truly sends chills down your spine is not a single product, but the simultaneous execution of all these capabilities.
Full modal understanding, full modal generation, and a 24/7 online agent—these three pieces of the puzzle were all put together by Google in one night. Omni turns a sentence into a world without needing human-provided materials; 93 agents built an operating system from scratch without a single line of code from humans; Spark works 24/7 for you without humans needing to open an app.
When AI no longer needs human 'feeding' but understands, decides, executes, and iterates on its own—the endpoint of this journey is called ASI (Artificial Super Intelligence).
No one can provide a precise timeline. But tonight’s Google I/O made everyone realize one thing: the barrier of 'technically impossible' on the road to superintelligence is no longer present. What remains is just the speed of engineering deployment. Six months ago, we were debating whether AGI was a bubble. Six months later, Google is already using agents to write operating systems. The acceleration in this industry has surpassed the range perceivable by human intuition.
Reference materials:
https://youtu.be/wYSncx9zLIU
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/
https://antigravity.google/blog/introducing-google-antigravity-2-0
https://antigravity.google/blog/google-io-2026-feature-deep-dive