Google’s AI endgame is here… everything you missed at I/O 2026

[00:00] Yesterday, Google I/O wrapped, and I was

[00:02] able to watch in person as Sundar and

[00:04] Demis laid out an ambitious vision for

[00:06] the future of software. And apparently,

[00:08] that future is Gemini hiding inside of

[00:10] every product like the microplastics in

[00:12] your bloodstream. But the road map is

[00:14] basically take Gemini, append a noun to

[00:16] it, and ship it. Gemini Spark, Gemini

[00:18] Omni, Gemini Flow, and the list goes on.

[00:21] But they're calling it the agentic

[00:23] Gemini era. The search is now an AI

[00:25] agent, Gmail is an AI agent, Android is

[00:27] an AI agent, your glasses are an AI

[00:29] agent. And as I watched the keynote, I

[00:31] realized something. That Google is no

[00:33] longer trying to organize the world's

[00:34] information with blue hyperlinks,

[00:36] because search engines are now an

[00:38] archaic technology. Instead, Google is

[00:40] trying to become the interface to

[00:42] reality itself before Anthropic and

[00:44] OpenAI create better realities. But

[00:46] luckily, Google I/O wasn't all about AI.

[00:49] I didn't see any updates about Angular,

[00:50] but I did come across a new awesome web

[00:52] API that every web developer should know

[00:54] about. In today's video, we'll break

[00:56] down everything you missed at Google

[00:57] I/O. It is May 22nd, 2026, and you're

[01:00] watching The Code Report. Whether you

[01:02] love it or hate it, one thing is

[01:03] undeniably impressive about Google, and

[01:05] that's its ability to scale. Not only is

[01:07] it serving its core products to billions

[01:09] of daily active users, but in the last 2

[01:11] years, they've gone from serving 9.7

[01:14] trillion tokens per month to a

[01:15] staggering 3.2 quadrillion tokens per

[01:18] month. And that number is going to

[01:19] continue accelerating. In addition,

[01:21] Alphabet's capital expenditures have

[01:23] exploded, building new infrastructure to

[01:25] support all these stupid AI images you

[01:27] guys create with nano banana. You ever

[01:30] see a pug dressed like an accountant?

[01:32] No.

[01:33] You want to? Uh

[01:35] One thing that makes this massive scale

[01:37] possible is Google's TPU chip, or Tensor

[01:39] Processing Unit. I remember being amazed

[01:42] seeing a TPU at my first Google I/O back

[01:44] in 2018. But this week, they announced

[01:46] they're splitting these chips into two

[01:48] distinct jobs, the training and

[01:50] inference with the TPU-T and TPU-I. In

[01:53] other words, Google now has one chip

[01:55] that's optimized to teach a robot how to

[01:57] think, and another chip that's optimized

[01:59] for it to hallucinate search results on

[02:00] a global scale. The headline

[02:02] announcement at Google I/O though was

[02:04] Gemini Omni, a model that takes any

[02:06] input like text, video, and sound and

[02:08] produces any output. Demis Hassabis, who

[02:11] might be the smartest guy at Google,

[02:12] appears to be fully world model pilled

[02:14] because models like this don't just

[02:16] generate pixels anymore. They understand

[02:18] language physics motion and

[02:20] everything else in your world just well

[02:22] enough to simulate reality on demand.

[02:24] But along with this new model comes an

[02:26] entirely new design system for the

[02:27] Gemini app called Neural Expressive. At

[02:30] first glance, the UI looks like a simple

[02:32] glow up with new icons and better

[02:33] gradients. But what's unique about it is

[02:35] that it's optimized for generating UI

[02:37] elements on demand, like diagrams,

[02:40] timelines, and even mini apps that

[02:41] didn't exist before your prompt. Now,

[02:43] when it comes to Google's core large

[02:45] language models, they released Gemini

[02:47] Flash 3.5, which is not the big brain

[02:49] model, but the fast model. According to

[02:51] the trust me bro benchmarks, it performs

[02:53] nearly on par with Opus 4.7 and GPT-5.5,

[02:57] but runs at a much faster speed. Like if

[02:59] we look at this trust me bro diagram, we

[03:01] see that Flash is entirely in a quadrant

[03:03] of its own in terms of speed and

[03:05] intelligence. However, it's important to

[03:07] remember that this is not their top-tier

[03:09] model. The Gemini 3.5 Pro is still under

[03:11] wraps and not expected to release until

[03:13] later this summer, which was very

[03:15] disappointing to a lot of people on the

[03:16] internet. Speaking of disappointment

[03:18] though, not everybody was happy with the

[03:20] new direction of Google's anti-gravity

[03:22] IDE. Anti-gravity was formerly known as

[03:24] Windserve and was code for AI coding

[03:27] just like Cursor. And once again,

[03:29] following in the footsteps of Cursor,

[03:30] its latest version looks like an OpenAI

[03:32] Codex clone that's more focused on

[03:34] managing agents than writing code. Old

[03:36] school programmers might not be happy

[03:38] with this change, but the live demo was

[03:40] pretty badass. They used the tool to

[03:41] build a complete operating system from

[03:43] scratch, which took like 12 hours and

[03:45] billions of tokens. But then, they tried

[03:47] to play Doom on it and it failed due to

[03:49] missing drivers. However, live on stage,

[03:52] they had Gemini code up those drivers

[03:54] and within a few seconds, Doom was up

[03:55] and running. The most impressive part

[03:57] was just the sheer speed at which this

[03:59] thing could spit out tokens. But, the

[04:00] speed is not the only thing increasing.

[04:02] But, the price of Gemini 3.5 Flash is

[04:04] three times more than the previous

[04:06] version and 30 times more than Gemini

[04:08] 1.5 Flash. It's still a lot cheaper than

[04:10] Claude, but not nearly as cheap as it

[04:12] used to be. Almost everything at IO

[04:13] involved AI in one way or another. But,

[04:15] if you're a web developer, one cool

[04:17] thing you should know about in Chrome is

[04:19] the HTML on Canvas API, which as the

[04:21] name implies, allows you to use HTML

[04:24] elements directly in a canvas now.

[04:26] >> Awesome. Native HTML elements rendered

[04:29] into the canvas.

[04:31] Woo!

[04:33] That means you can build highly

[04:34] interactive UIs where you control every

[04:36] pixel with tools like WebGL and WebGPU,

[04:39] while simultaneously using HTML for your

[04:42] more basic UI elements. The only

[04:44] question is which AI coding model should

[04:46] you use to work with this API? Well,

[04:48] that's why you need to know about

[04:49] Emergent, the sponsor of today's video.

[04:51] Everyone's switching between five

[04:52] different coding models these days, but

[04:54] we still need something to help us ship

[04:56] full stack applications that actually

[04:58] work. And that's exactly where Emergent

[04:59] can help. Right now, I'm using it to

[05:01] build a pull request review dashboard

[05:04] where I can paste in a GitHub link and

[05:06] get an AI summary of all the changes and

[05:08] risks per repo. You still start with a

[05:09] prompt, but instead of one LLM guessing

[05:12] how to build everything, Emergent spins

[05:14] up specialized agents to work on your

[05:16] app's front end, back end, database,

[05:18] testing, and deployment all in parallel.

[05:20] You also don't need to mess with any

[05:22] Superbase wiring or Express boilerplate,

[05:24] because that one prompt sets up your

[05:26] app's database, auth, and APIs. If

[05:28] you're really into self-torture, feel

[05:30] free to keep scaffolding this stuff by

[05:31] hand, or you could just describe the

[05:33] tool you want and let Emergent's agents

[05:35] swarm build it all for you. You try it

[05:37] out for free at the link below. This has

[05:39] been the Code Report. Thanks for

[05:40] watching and I will see you in the next

[05:42] one.