The Brief #26.42 · Gemini 4 Argon, Claude Haiku 5.5, and Firefox 158

Developers can finally use Gemini 4 Argon!

Meta’s Muse Wonder Full came third on Artificial Analysis’ Intelligence Index, Anthropic released Claude Haiku 5.5, and Firefox 158 is out with typed arithmetic in CSS.

Then there’s Core Web Vitals for single-page apps, what’s coming to browsers, more decision models, Intelligent UI in ChatGPT, Genkit Dart 1.0 for Flutter apps, and Vercel’s deployment storage charges.

Gemini 4 Argon and Muse Wonder

On October 13, Google opened Gemini 4 Argon to developers.

Developers can call it as gemini-4-argon on the paid tier of the Gemini API, and use it in Google AI Studio and Android Studio. Google AI Ultra subscribers can also try it in AI Studio and in Antigravity.

The Gemini API’s Free tier doesn’t include Argon. In the Gemini app, though, AI Pro and Ultra subscribers can use it.

The API price is $2 input / $10 output per million tokens through November 30, and $4 input / $20 output from December 1. That introductory price had no end date until now.

Meta released Muse Wonder and Muse Wonder Full on October 12. Muse Wonder is in preview on the Meta Model API as muse-wonder-1.0. It has a context window of 1M tokens and costs $4 input / $20 output per million tokens.

Muse Wonder Full is only in Muse Code, Meta’s coding agent for the terminal. It comes with the Power Usage and Maximum Usage subscriptions. Meta doesn’t sell it per token.

Artificial Analysis scored Muse Wonder Full at 56 and Muse Wonder at 52, both at max effort. That’s up from 48 for Meta’s Muse Spark 1.3 only six weeks earlier. Muse Wonder Full’s 55.8 puts it third, just behind Claude Sonnet 5.5’s 56.0, and Muse Wonder is eighth, at $4.52 per task.

Claude Sonnet 5.5’s score didn’t change, but its cost per task went down from $7.67 to $5.46. Anthropic halved Sonnet 5.5’s cache-read price on October 7, from $0.20 to $0.10 per million tokens.

Claude Haiku 5.5

Anthropic released Claude Haiku 5.5 on October 7. Haiku is the smallest and cheapest Claude line, made for quick jobs you run many times, like summarizing text or sorting messages into categories.

Haiku 5.5 costs $0.10 input / $0.50 output per million tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5’s $1 input / $5 output. Once a prompt goes over 100,000 tokens, cached parts included, you pay the higher prices for the whole request, $0.50 input / $2.50 output. Haiku 5.5 matches GPT-6 Luna’s price for prompts under 100,000 tokens, but Luna stays at that price up to 272,000 tokens.

It’s the first Haiku with an effort setting, from low to max, for trading the quality of an answer against speed and cost.

At max effort, Artificial Analysis scored Haiku 5.5 at 43, 26 points more than Haiku 4.5. It used about 162,000 output tokens per task, around three times as many as GPT-6 Luna.

Along with Haiku 5.5, Anthropic started including monthly API credits in the Claude Max and Team plans. Max 5x comes with $100 a month and Max 20x with $200. Team plans get $20 per Standard seat and $100 per Premium seat, in one shared pool of up to $500.

The credits cover the Claude API, including Managed Agents and the Agent SDK, but not Claude Code or Claude on other clouds like Google Cloud and Amazon Bedrock. They don’t roll over to the next month. To claim them, you link one organization from the Claude Console (the dashboard for API keys and billing) in your plan’s billing settings, and you have to contact support to change it later.

Firefox 158

Firefox 158 came out on October 13 with JPEG XL on by default.

On the CSS side, it adds typed arithmetic to calc().

calc() could already multiply or divide a length by a plain number, like calc( 100vw / 2 ), but dividing a length by another length didn’t work. Typed arithmetic allows it, and the units cancel out. 100cqw is the full width of the nearest element with container-type set, and if that element is 480px wide, calc( 100cqw / 1px ) gives you the plain number 480.

You can use that number in properties that don’t take a length, like font-weight, or multiply it by 1deg to turn it into an angle. The CSS below turns a sale card into a container and makes its heading bolder as the card gets wider.

.sale-card {
	container-type: inline-size;
}

.sale-card h3 {
	/* The card's width as a plain number, like 240 */
	--card-width: calc( 100cqw / 1px );

	/* 400 at 160px or narrower, 800 from 320px */
	font-weight: clamp( 400, var( --card-width ) * 2.5, 800 );
}

A browser without typed arithmetic can’t compute that font-weight value and shows the heading at the card’s regular weight.

Chrome added typed arithmetic in version 140 and Safari in version 26, which means it now works in all three browsers.

For notifications, there’s a new navigate option. When a user clicks the notification, the browser opens the URL you passed in navigate. Without it, you’d handle the notificationclick event in your service worker and open the page yourself. It has worked in Safari since version 18.4, but not in Chrome.

In version 158, the browser suggests saved passkeys (sign-ins with a fingerprint or the device’s PIN instead of a password) on login fields with autocomplete="webauthn" and no other value. Before, the attribute had to include another value, like autocomplete="username webauthn".

Core Web Vitals for single-page apps

On October 13, Google started publishing experimental Core Web Vitals for single-page apps in the Chrome UX Report, or CrUX. CrUX is the public data Chrome collects on how real users experience about 18 million websites.

The Core Web Vitals are three of its measurements. LCP measures how fast the main content shows up, and INP how fast the page responds to an interaction, like a click or a key press. CLS measures how much the layout jumps around.

A single-page app loads one HTML page from the server and updates the content and the address with JavaScript from there. Those page changes are called soft navigations.

The stable CrUX numbers count a visit to a single-page app as one long page view. Only the first load gets an LCP, and the clicks and layout shifts from the whole visit count against that first page. The experimental versions measure every soft navigation as a separate page view, under the URL the user went to.

Google expects little or no change for the roughly 70% of sites in CrUX that have almost no soft navigations. But 35% of the top 1,000 sites on Android and 42% on desktop use soft navigations for more than half of their page views.

Those sites mostly score better with the experimental numbers. Among sites where 75% to 100% of page views are soft navigations, 87% have a good LCP on Android in the experimental data, against 26% in the stable data. The later pages load faster because they reuse code and files the browser downloaded for the first one.

For sites with mostly soft navigations, the INP pass rate on Android roughly doubles, but single-page apps still score lower on INP than regular multi-page sites. Google says heavy JavaScript during a soft navigation can cause INP problems.

Coming to browsers

Safari Technology Preview 254 came out on October 8 with interpolate-size and calc-size(), the two CSS features for animating to height: auto. It’s the early version of Apple’s browser that developers use to try features before they ship.

An accordion can animate from height: 0 to height: auto, the height of its content, once you set interpolate-size: allow-keywords on :root. To do math on those sizes, you use calc-size(), where calc-size( auto, size + 2rem ) is the height of the content plus 2rem. In browsers without interpolate-size, the accordion opens right away, with no animation.

Chrome added both in version 129. Firefox doesn’t have them. The preview also adds symbols(), the function for custom list markers that Chrome added in version 155 and Firefox has supported since version 35.

The HTML standard started allowing a <style> element in the body on October 9, as long as it’s the first child of its parent, it has no @import, and its rules are inside an @scope block with no selector. That block limits the styles to the <style> element’s parent and everything inside it.

<article>
	<style>
		@scope {
			:scope { border: 1px solid; }
			img { max-width: 100%; }
		}
	</style>
	<img src="desk-lamp.jpg" alt="A brass desk lamp with a green shade">
	<h3>Desk lamp</h3>
</article>

The :scope selector matches the <article> itself. The standard requires that first-child position so a <style> in the body can’t restyle content the browser already drew, other than its parent.

Browsers apply a <style> element in the body anyway, and every major browser supports @scope. Components can keep their styles next to their markup this way today, in valid HTML.

More decision models

Microsoft released Microsoft-Decision-1 on October 9, in Microsoft Foundry and through OpenRouter.

I explained decision models three weeks ago with TypeSafe AI’s Jev. Like Jev, Microsoft-Decision-1 returns one of the answers you allow, with a probability for each, instead of writing text.

It has the same price as Jev, $0.042 input / $0 output per million tokens. Besides yes-or-no and multiple-choice questions, it can grade an AI’s answer or an agent’s next step against criteria you write. Microsoft says that in its tests across 36 benchmarks, it was the most accurate model and answered about 35 times faster than GPT-6 Sol.

Artificial Analysis started scoring decision models on October 13, combining four tests into one index. Jev 1.13 led the decision models with 58, and Microsoft-Decision-1 tied for second with 54. Microsoft’s model was cheaper to run, at $31 per million tasks against Jev’s $42.

Every decision model in the index was overconfident, which means their answers were right less often than the probability they gave them. That happened mostly on the knowledge and reasoning questions, and much less on the classification and tool calling tests. If your code acts on answers above a certain probability, more of them will be wrong than you’d expect from that number.

Claude Haiku 5.5 scored 61 on those tests at low effort, 3 points more than Jev, but it was about 11 times slower and cost 4.4 times as much.

On October 7, Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M. d1-3B takes text and images. The smaller d1-omni-600M takes text with an image or an audio clip, but Liquid calls it experimental. Liquid measured d1-3B answering a question in 30 ms on Apple’s M5 Pro chip and in 8 ms on an NVIDIA RTX 4090.

With the weights on Hugging Face, you can run them on your servers or devices, and the data you classify never goes to an outside API.

Cloudflare released Clef-omni on October 9 as the first of its Clef decision models that takes audio and video along with text and images. It costs $0.15 per million input tokens.

Cloudflare also lowered Clef-flash’s price from $0.09 to $0.038 per million input tokens, below Jev’s $0.042. To get there, it shrank the context window of the hosted Clef-flash from 64,000 to 24,000 tokens. According to Cloudflare, 0.24% of requests were longer than that.

TypeSafe AI raised $870 million on October 9, in a Series A funding round led by Andreessen Horowitz. The round values the company at $7.5 billion.

Intelligent UI in ChatGPT

OpenAI brought GPT-6 with Intelligent UI to ChatGPT on October 7 for the Plus, Pro, Business, and Enterprise plans, and on October 8 for the Free and Go plans. With Intelligent UI, ChatGPT can answer with interactive pieces like charts and forms, or even a small tool, such as a calculator that shows how your savings could grow.

OpenAI trained GPT-6 to add visuals and interactive parts to an answer if they help, and to stick to plain text otherwise.

Genkit Dart 1.0

On October 8, Google released Genkit Dart 1.0, the first stable version of its open-source framework for building AI features and agents in Dart. Genkit was already available for TypeScript, Go, and Python, and the Dart version had been in preview since March.

Genkit gives you one API for Gemini, Claude, OpenAI’s models, and any model with an OpenAI-compatible API. Each model provider is a plugin. When you switch from one model to another, the rest of your code doesn’t change.

import 'package:genkit/genkit.dart';
import 'package:genkit_anthropic/genkit_anthropic.dart';
import 'package:genkit_google_genai/genkit_google_genai.dart';

void main() async {
	final ai = Genkit( plugins: [ googleAI(), anthropic() ] );
	final prompt = 'Suggest a weekend getaway from San Francisco.';

	final fromGemini = await ai.generate( model: googleAI.gemini( 'gemini-flash-latest' ), prompt: prompt );
	final fromClaude = await ai.generate( model: anthropic.model( 'claude-sonnet-5-5' ), prompt: prompt );

	print( fromGemini.text );
	print( fromClaude.text );
}

The 1.0 packages are on pub.dev, and dart pub add genkit adds the main one to your project.

Vercel deployment storage

Vercel said on October 9 that every team on its Pro plan will pay for deployment storage. New teams started paying for it on August 21. Existing teams get the start date by email.

When you publish your site to Vercel, it keeps a copy of that version, called a deployment. That copy lets you open an old preview or roll production back to it if a release breaks something. The build output and static files count as Deployment Storage, the function bundles count as Functions Storage, and both cost $0.10 per GB-month. One GB stored for a whole 30-day month is 1 GB-month.

Starting October 23, Vercel will also delete deployments older than 30 days, unless you choose a longer retention period by then. Teams whose retention is already 30 days or shorter keep their settings. You can’t roll back to a deleted deployment.

Vercel’s reason is that coding agents made teams deploy more often, which means more stored history. New Pro teams start with a 30-day retention period.