The Brief #26.46 · Gemini 4 Carbon for developers, Claude Fable 5.2, and Modern Web Guidance

Want more of these in your Google results?
Google’s Gemini 4 Carbon is open to developers, and Claude Fable 5.2 got the highest score on Artificial Analysis’ Intelligence Index.
The Chrome team released Modern Web Guidance 1.0 to teach coding agents how to use modern browser features.
Adaptive Dark Mode in the Safari 27.3 beta can darken light websites that don’t have a dark theme.
Besides that, you can now use speech recognition in Firefox 160, make 15-second videos with Gemini Omni 1.2 Flash, and get SVG images from OpenAI’s image API.
Gemini 4 Carbon and Claude Fable 5.2
Google opened Gemini 4 Carbon to developers on November 10. Until that day, only some security teams could use it.
You can call it on the paid tier of the Gemini API or straight from a web app through Firebase AI Logic (that needs the pay-as-you-go Blaze plan), and use it in Antigravity if you have a Google AI Ultra subscription.
Carbon costs $2 input / $10 output per million tokens through November 30. The regular price starts on December 1, at $4 input / $20 output per million tokens.
Google says Carbon spends more tokens than Gemini 4 Argon on each answer and takes longer to respond, because it tries more ideas and checks its work first. That’s why Google still suggests Argon for most software work.
Anthropic released Claude Fable 5.2 on November 9. The price went down to $8 input / $40 output per million tokens, from $10 input / $50 output for Fable 5.1. According to Anthropic, Fable 5.2 also writes its answers up to 30% faster.
Similarly to Google, Anthropic recommends starting with Claude Opus 5.5 for most coding and knowledge work, and moving up to Fable for the most ambitious tasks.
At max effort, Fable 5.2 got 65 on Artificial Analysis’ Intelligence Index, the highest score on the index, but at the extremely high cost of $26.54 a task. It got 62 at high effort, for $9.08 a task.
Gemini 4 Carbon scored 60 at $7.76 a task. No other model reaches 60 for less. Artificial Analysis calculated the $7.76 with the regular price, which means a task costs less while the introductory price lasts.
| # | Model | Intelligence Index | Cost per Index task |
|---|---|---|---|
| 1 | Claude Fable 5.2 (max)New | 65 | $26.54 |
| 2 | Claude Fable 5.2 (high)New | 62 | $9.08 |
| 3 | Gemini 4 Carbon (high) | 60 | $7.76 |
| 4 | Claude Opus 5.5 (max) | 59 | $10.04 |
| 5 | GPT-6 Astra (max) | 57 | $6.24 |
| 6 | GPT-6.1 Sol (max) | 55 | $1.27 |
| 7 | Claude Sonnet 5.5 (max) | 55 | $10.02 |
| 8 | Claude Opus 5.5 (high) | 52 | $3.46 |
| 9 | GPT-6.1 Terra (max) | 51 | $0.96 |
| 10 | Gemini 4 Argon (high) | 50 | $9.95 |
| 15 | Muse Ember (max)New | 36 | $0.14 |
| # | Model | Decision Index | Cost per million tasks |
|---|---|---|---|
| 1 | Drex Max v1.1 | 73 | $102.68 |
| 2 | GLiDE v1.2 | 73 | $780.40 |
| 3 | Mammoth 27BNew | 71 | $263.57 |
| 4 | d1-ultra | 71 | $133.04 |
| 5 | Decider v1.2 | 70 | $68.46 |
| 6 | SubQ Snap 1.1 | 68 | $136.90 |
| 7 | Elephant 9B | 66 | $92.25 |
| 8 | Jev 1.14 | 66 | $156.23 |
| 9 | Microsoft-Decision-1 | 62 | $142.56 |
| 10 | Mercury Decide | 62 | $69.53 |
| 11 | Celeris-1.1 DecisionNew | 62 | $13.67 |
| Chrome’s Decisions API (Gemma 4 on CPU) | Not scored | $0 |
Meta released Muse Ember on November 9. It’s a small model for the jobs an agent repeats all day, like calling tools or pulling data out of documents. It costs $0.10 input / $0.50 output per million tokens and has a context window of 1M tokens.
Artificial Analysis released version 2 of its Decision Index on November 8, with new tests and a different way of scoring them. Don’t compare the scores or costs of decision models with last week’s, because the two versions measure different things.
On Monday I published an article about decision models and Chrome’s Decisions API, in case you want to see what you can build with them.
Modern Web Guidance
The Chrome team released Modern Web Guidance 1.0 on November 10!
It’s a set of guides for coding agents like Antigravity or Claude Code. The guides teach an agent the right way to use the features browsers have today.
According to the team, AI models are very good at making something work with JavaScript, even where HTML or CSS can do it with much less code. In Google’s tests, models got things wrong in areas like keyboard focus, form validation, light and dark themes, and layout shift (content that jumps around while the page loads).
The team runs those tests every night with and without the guides and publishes the results. Agents passed 53% of the checks without the guides and 92% with them in the runs from November 9.
You install Modern Web Guidance as a skill. Run this command from your project’s folder:
sh
npx modern-web-guidance@latest installThe project’s experts reviewed every guide for this release, with an extra accessibility review for the ones about user interfaces. The guides come with data from Baseline so the agent can check which features work in the browsers you support.
The default target is Baseline Widely available, features that have worked in every major browser for at least 30 months. You can set a different one in the instructions file your agent uses, like AGENTS.md.
Version 1.0 adds a workflow for existing projects.
To start it, send your agent /modern-web-guidance Modernize the code in this directory.
A scanner checks the project’s frontend code and its package.json for old patterns that are safe to replace.
For example, it finds JavaScript that makes a <textarea> grow as you type.
CSS can do that with field-sizing: content.
Then the workflow guides the agent through each change.
The scan runs on your computer and doesn’t use any tokens. The team started with a small set of patterns and plans to add more.
So, try that prompt on a project that’s a few years old.
Google says Firefox and Edge are joining Chrome to work on the project.
Safari’s Adaptive Dark Mode
Apple’s WebKit team announced Adaptive Dark Mode on November 3. It’s a feature in the Safari 27.3 beta that shows light websites in dark colors even if they don’t have a dark theme.
People have to turn Adaptive Dark Mode on themselves. The browser offers them a preview if their system is in dark mode and they open a bright website.
WebKit says it built the feature because about 40% of page loads come from people who use dark mode, but only about 10% are on pages that declare a dark theme. Both numbers are from Chrome’s usage statistics.
The browser leaves a page alone if it already looks dark.
Light pages get dark backgrounds and light text, but most images and anything drawn on a <canvas> don’t change.
WebKit suggests SVG for logos and icons, because the browser can change an SVG’s colors if they’d be hard to see on a dark background.
A page that declares support for dark mode isn’t changed at all.
The declaration is <meta name="color-scheme" content="light dark"> in the HTML or color-scheme: light dark in the CSS.
The WebKit team says not to add it later with JavaScript.
And you have to write the dark colors yourself, for example with light-dark(), a CSS function that takes a color for light mode and a color for dark mode.
light-dark() works in Chrome, Firefox, and Safari.
html
<!-- Put this in the head. The page supports light mode and dark mode. -->
<meta name="color-scheme" content="light dark">
<style>
/* The same declaration in CSS. */
:root {
color-scheme: light dark;
}
/* The first color is for light mode, the second for dark mode. */
body {
background-color: light-dark( #ffffff, #1f1d1b );
color: light-dark( #24211e, #f2eee8 );
}
</style>There’s more about light-dark() in my August article, Let CSS Pick Your Text Color.
WebKit says a page that has to stay light can declare light only instead of light dark, but Safari 27.3 beta 1 doesn’t support the only keyword yet.
Your JavaScript can’t detect Adaptive Dark Mode either, because code that checks an element’s color gets the color you wrote.
Firefox 160
Firefox 160 came out on November 10 with speech recognition for web pages.
A page can turn what a person says into text with SpeechRecognition, part of the Web Speech API.
You could use it for a button that lets people dictate a message into a form.
Firefox recognizes 40 languages and does it on the user’s computer, without sending the audio to a server. The browser has to download a language pack for each language once. After that, recognition works offline.
Chrome has been able to recognize speech on the user’s computer since version 139. That means one piece of code works in both browsers on desktop.
Google also announced speech tools on November 10, including two tasks for its MediaPipe library that work in the browser.
The tasks are Audio Transcriber for turning speech into text on the user’s computer and Text-to-Speech for reading text out loud.
Both work with open-weight models, like Parakeet for recognition and Kokoro for speech, and come in the @mediapipe/tasks-audio package on npm.
Gemini Omni 1.2 Flash and SVG images
Gemini Omni 1.2 Flash has been generally available in the Gemini API and Gemini for business since November 5. Omni makes short videos with sound from text or images, and edits existing ones.
A clip can now be up to 15 seconds long, up from 10 in Omni 1.1. The price is 30% lower too, at about $0.07 per second of 720p video. A 15-second clip comes to about $1.04.
You can edit a clip by describing the change, like a different background or lighting, instead of generating the whole video again.
Also, OpenAI put out two image models on November 4, gpt-image-2.6-sunburst for detailed work and gpt-image-2.6-flare for speed.
Both can return an SVG or an animated GIF instead of a PNG.
The API makes a regular image first and then draws the SVG from it, which means you pay for that image and for the tokens it takes to draw the SVG.
AlloyDB Express
Google Cloud introduced AlloyDB Express on November 4, in Preview (you can use it, but it can still change). It’s a small PostgreSQL database for tests and prototypes, and it can be ready about 2 seconds after you create it.
The database pauses after 15 minutes without activity, which means you don’t pay for compute while nobody uses it. It takes about 20 seconds to resume.
Keep the limits in mind, though. You get 1 or 2 virtual CPUs and up to 20 GB of storage, and there are no backups. When the project grows, you can upgrade it to a full AlloyDB database and keep the data and the connection string.
There’s a free tier for the first 12 months, and it’s available in seven regions.
Want more of these in your Google results?