Agustin Barrientos
All posts

The Brief #26.47 · Gemini 4 Hydrogen, GPT-6.2, and Chrome 158

Want more of these in your Google results?

Google released Gemini 4 Hydrogen for everyday coding, and a cloud API for decision models called the Gemini Decisions API.

OpenAI released GPT-6.2 Astra for the hardest projects and GPT-6.2 Sol for everyday work.

On the web side, Chrome 158 lets you style a video player with CSS depending on whether it’s playing, paused or muted.

There’s also A2UI 1.0, Google’s protocol for agents that answer with your own components, and Flutter 3.50, which can now run as WebAssembly in Safari and Firefox.

Plus, Firebase Crashlytics for websites, the new HTML switch, and ChatGPT Wallet.

Gemini 4 Hydrogen, the Gemini Decisions API, and GPT-6.2

Google released Gemini 4 Hydrogen on November 12, a fast model for everyday coding and agent tasks. If you’re using Gemini 3.8 Flash or an older Gemini 3 or 2.5 model, Google recommends moving to Hydrogen. For classifying and routing lots of requests, though, they still suggest Gemini 3.5 Flash-Lite.

Antigravity got a new option too, called Orchestrator. Gemini 4 Carbon or Argon plans the task while Hydrogen does the work, like editing the code and running the tests.

The day before, Google released the Gemini Decisions API, with a new model called Gemini Decision 1. You send it up to 1,000 questions in one request, about text, images, audio, video or PDFs, and it gives you back an answer for each one with its probability. A question can be yes or no, multiple choice, or a score on a scale you choose.

According to Google, a request with up to 16K tokens takes a median of 150 to 250 ms. And it’s really cheap. You pay $0.10 per million input tokens and nothing for the output, or nothing at all on the free tier (up to 10,000 requests per day!).

I added it to my article about decision models on November 11, with code and the results of my own tests, in case you want to try it.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. According to OpenAI, both get better results on hard tasks while using fewer tokens.

In the API, Astra costs $8 input / $40 output per million tokens, which is 20% less than GPT-6 Astra. Sol keeps the same price as GPT-6.1 Sol, $2 input / $10 output. You can also use both in ChatGPT and Codex.

So, how does Hydrogen compare? Hydrogen got 54 on Artificial Analysis’ Intelligence Index at xhigh, its new highest thinking level, for $3.16 a task. At high, which is the default, it got 51 for $1.87, one point more than Gemini 4 Argon.

By the way, Artificial Analysis now uses the introductory price to calculate what Argon and Carbon cost per task.

On the decision side, Gemini Decision 1 scored 66 on Artificial Analysis’ Decision Index. The top spot went to Liquid AI’s d1.2-ultra, released on November 16, with 75. Perplexity also released Decider v1.3 and Decider Flash v1.3 with open weights on November 12, a day after Subquadratic put out SubQ Snap 1.2.

Table Intelligence Index
#ModelIntelligence IndexCost per Index task
1Claude Fable 5.2 (max)65$26.54
2GPT-6.2 Astra (max)New62$4.00
3Claude Fable 5.2 (high)62$9.08
4Gemini 4 Carbon (high)60$5.48
5Claude Opus 5.5 (max)59$10.04
6GPT-6.2 Sol (max)New57$1.01
7Claude Sonnet 5.5 (max)55$10.02
8Gemini 4 Hydrogen (xhigh)New54$3.16
9Claude Opus 5.5 (high)52$3.46
10GPT-6.1 Terra (max)51$0.96
11Gemini 4 Argon (high)50$4.97
12Muse Wonder Full (max)50No API price
13Kimi K3.1 ProOpen weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave it46$3.68
14Muse Wonder (max)44$6.95
15Qwen4 MaxNew43$3.27
16Claude Haiku 5.5 (max)36$0.41
17Muse Ember (max)36$0.14
Table Decision Index
#ModelDecision IndexCost per million tasks
1d1.2-ultraNew75$133.04
2Drex Max v1.173$102.68
3SubQ Snap 1.2New73$136.90
4GLiDE v1.2Open weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave it73$780.40
5Mammoth 27BOpen weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave it71$263.57
6Decider v1.3Open weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave itNew70$68.04
7d1.2New69$63.35
8Elephant 9BOpen weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave it66$92.25
9Jev 1.1466$156.23
10Gemini Decision 1New66$345.41
11Decider Flash v1.3Open weights: the model files are downloadable, so you can run it on hardware you choose and your prompts never have to leave itNew64$34.02
Chrome’s Decisions API (Gemma 4 on CPU)Not scored$0

A2UI 1.0

Google released A2UI 1.0 on November 11. A2UI is an open protocol that lets an AI agent answer with an interface, like a form, a card or a chart, instead of text.

The agent can only use components you registered in your app, like your own buttons and date pickers. A renderer in your app draws them, so the answer looks like the rest of your interface. There are open-source renderers for Angular, Flutter, Lit, React, and SwiftUI.

Agents can write those screens in A2UI Express, the short format I covered in Brief 26.44, with one line per component.

This is the first stable version of the protocol. Don’t let the version numbers on npm confuse you, though. The web renderers have their own, like 0.15.0 for @a2ui/lit and @a2ui/react.

With 1.0, the agent and your app can call functions on each other. In Google’s example, a reservation picker checks availability right away.

Accessibility is part of the spec now too. Renderers have to pass each component’s label and description to screen readers, which on the web means aria-label and aria-describedby. When the agent doesn’t send a label, the renderer uses the component’s visible text instead. Your own components have to follow the same rules.

The easiest way to see it working is the quickstart. You paste a prompt into a coding agent like Antigravity. It runs three commands to set up a small trip planner, then explains how A2UI works while the planner is open.

Next on the roadmap is official support for decision models in A2UI’s libraries for building agents.

Chrome 158

Chrome 158 came out on November 17 with new CSS pseudo-classes for video and audio.

The new ones are :playing, :paused, :seeking, :buffering, :stalled, :muted, and :volume-locked. They work in Chrome from version 158, Firefox from 150, and Safari from 15.4. Chrome’s 150 and 152 betas already had them, but 158 is the first stable version with them.

Browser supportMedia pseudo-classes20% of users
Chrome158
Edge158
Firefox150
Safari15.4
Safari iOS15.4
Chrome Android158

Browser data updated November 12, 2026

That means that if you build your own video player, CSS can show its play button, a loading note or a muted badge, without JavaScript listening to the video’s events. In the CSS below, the play button disappears while the video plays, and the badge shows up while it’s muted:

css
/* The play button and the badge come after the video, inside the same wrapper. */

/* Hide the big play button while the video plays. */
video:playing ~ .play-button {
	display: none;
}

/* Show the badge only while the video is muted. */
.muted-badge {
	display: none;
}

video:muted ~ .muted-badge {
	display: block;
}

:buffering is the one for a loading note. It matches while the video is playing but stuck, waiting for more data. If it stays stuck for too long, it matches :stalled too.

CSS can’t do everything here, though. You still need JavaScript to actually start the video when someone presses the play button, and for things like showing how much of it has played. These are all for styling, not controlling.

HTML switch and Safari 27.3 beta 2

HTML got a switch control on November 16, the on/off toggle you see in phone settings. You make one by adding the switch attribute to a checkbox:

html
<!-- Sends alerts=on with the form while it's turned on -->
<label><input type="checkbox" name="alerts" switch> Email alerts</label>

Only Safari draws it as a switch so far, since version 17.4. Chrome and Firefox have open bugs for it. But you don’t have to wait, because a browser without support shows a normal checkbox, and the form sends alerts=on either way.

Safari 27.3 beta 2 also came out on November 16, with support for color-scheme: only light. With that line, a page or a single element stays light even when Adaptive Dark Mode (Safari’s setting that darkens light websites) is on. Chrome’s Auto Dark Theme on Android already respects the same opt-out.

The same day, Apple’s WebKit team showed a preview of spatial CSS, a proposal for putting 3D models in a page and placing them with regular CSS:

html
<!-- The 3D models inside a portal share one space -->
<div style="spatial: portal;">
	<model id="board" src="board.usdz"></model>
	<model id="star" src="star.usdz"></model>
</div>

<style>
	/* Move the star to the right of the board. */
	#star {
		transform: translateX( 50cm );
	}
</style>

It’s very early. The CSS Working Group published it that day as an Editor’s Draft, a first version of a spec that can still change at any moment. You can only try it on the visionOS 27.3 beta, on Apple’s Vision Pro headset or in the visionOS simulator you can install from Xcode, after turning on ⁠CSS Portals⁠ in Safari’s advanced settings.

Flutter 3.50

Flutter 3.50 and Dart 3.14 came out on November 11. Flutter is Google’s framework for building phone, desktop and web apps from one Dart codebase.

The biggest change for the web is that Safari and Firefox can now run your Flutter app as WebAssembly. Flutter could already build web apps as WebAssembly with flutter build web --wasm, but only Chrome and Edge 137 or later used that build. The other browsers got the JavaScript one.

Now you can turn it on for Safari 26.2 and Firefox 147 or later too. In Google’s benchmarks with heavy widgets, WebAssembly ran at more than twice the frame rate of JavaScript in those two browsers. The Flutter team plans to turn it on by default for them in a future stable release. For now, you opt in with wasmAllowList in web/flutter_bootstrap.js:

js
// web/flutter_bootstrap.js
// The build replaces these two tokens with Flutter's loader and settings.
{{flutter_js}}
{{flutter_build_config}}

// Run the WebAssembly build in Chrome and Edge (blink),
// Safari and the other iOS browsers (webkit), and Firefox (gecko).
_flutter.loader.load( {
	config: {
		wasmAllowList: {
			blink: true,
			webkit: true,
			gecko: true,
		},
	},
} );

The Dart team also deprecated the old libraries for the browser and JavaScript, like dart:html, dart:js and package:js. They don’t work with WebAssembly and will be removed in a future release. Their replacements, package:web and dart:js_interop, work in both builds. To check whether your app or one of its packages still imports an old one, build with flutter build web --no-deprecated-js-interop. The build fails if it does.

Last, JNIgen (for Java and Kotlin) and FFIgen (for Objective-C) are stable now, so Flutter apps can call Android and Apple APIs directly from Dart.

Crashlytics for websites

When your site throws an error in someone’s browser, you don’t see it unless that person tells you. Firebase Crashlytics collects those errors and turns them into reports in the Firebase console.

Since November 12, it works with websites too, wherever you host them. Your site only needs a Firebase project and the Firebase SDK. Before, it was only for Android and Apple apps, Flutter and Unity, which means that if you have a mobile app and a website, all their errors end up in one dashboard.

With your source maps, the reports point to your original files and lines. Just keep in mind it’s still a Preview, and it needs the pay-as-you-go Blaze plan.

Setting it up is one call to getCrashlytics() where you initialize Firebase. This is based on Google’s setup guide:

js
import { initializeApp } from '@firebase/app';
import { getCrashlytics } from '@firebase/crashlytics';

// Your web app's settings, from Project settings in the Firebase console.
const firebaseConfig = {
	apiKey: 'API_KEY',
	authDomain: 'AUTH_DOMAIN',
	projectId: 'PROJECT_ID',
	storageBucket: 'STORAGE_BUCKET',
	messagingSenderId: 'MESSAGING_SENDER_ID',
	appId: 'FIREBASE_APP_ID',
};

const app = initializeApp( firebaseConfig );

// Report every error your code doesn't catch, with the app's version.
getCrashlytics( app, {
	appVersion: '2.4.0',
} );

There’s also provideCrashlytics() for Angular and a <FirebaseCrashlytics> component for React.

Gemini’s Deferred tier on Google Cloud

On November 16, the Deferred tier of Gemini Enterprise Agent Platform became generally available. Agent Platform is where you use Gemini on Google Cloud, and with the Deferred tier, a request costs 50% less if you can wait for the answer.

Google aims to finish 95% of them within 24 hours, without guaranteeing it. So it’s for work nobody is waiting on, like an agent that updates code overnight or goes through a pile of documents.

You ask for it with service_tier: "deferred" in the Interactions API, together with background: true and store: true, and it only works on the global endpoint. The list of models has Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.1 Pro and several Flash models, but not Gemini 4 Hydrogen.

ChatGPT Wallet

On November 13, OpenAI introduced ChatGPT Wallet, which lets an AI agent buy things on regular websites for you. You still approve every purchase yourself, with something like Face ID or your fingerprint. It’s starting in dots (OpenAI’s assistants that keep working on tasks for you), for ChatGPT Pro subscribers in the US first.

Instead of your card, the agent pays with a payment token, which means the store never sees your real card number.

If you sell online, OpenAI says Wallet works with your existing checkout pages and payment forms, without any Wallet-specific code. Your bot protection might block the agent, though. OpenAI suggests letting your CDN or firewall accept requests that ChatGPT signs with Web Bot Auth (a signature that proves the request comes from ChatGPT). Akamai, Cloudflare, HUMAN, and Vercel check those signatures automatically.

Want more of these in your Google results?