01Gemini 4 Carbon vs Fable 5.2: What This Test Is

The gemini 4 carbon vs fable 5.2 test is a head-to-head race between two brand new frontier models that launched one day apart. In this episode, Julian Goldie gave both models the same five real business builds, with the same prompt, the same start signal and a real clock running.

Anthropic released Claude Fable 5.2 first. One day later, Google said Gemini 4 Carbon was available. Both models are built for long jobs, and both can read up to a million tokens of context.

Each build was called a lap. Every lap was something a member of the AI Profit Boardroom asked about that week. At the end of each lap, both builds got the same four checks:

  • Speed: how long the model took to finish.
  • Works first try: does the app actually run when you open it?
  • Design: does it look like a finished product?
  • Follows the brief: did it build what was asked?

The final score came down to the very last lap. One model shipped a broken app and said it passed. The other lost three laps on design.

02Where to Get Each Model and What the Benchmarks Say

Fable 5.2 runs in the Claude apps and in Claude Code for everyone. Gemini 4 Carbon is rolling out in Antigravity and the Gemini app for Google AI Ultra subscribers, and to developers in AI Studio.

On paper, the two models are close. Here is what the published numbers in this episode show:

  • Artificial Analysis Intelligence Index: at the same high setting, Fable scores 62 and Carbon scores 60.
  • Frontier Suite: Anthropic's own table puts Fable ahead, 69.9 to 61.2.
  • Automation Bench: that same table has Carbon ahead, 53.9 to 51.4.

So each model wins some benchmarks and loses others. That is why Julian wanted to see what happens when you hand each one a real job instead of a test sheet.

Benchmarks tell you how a model does on fixed questions. They do not tell you if the app it builds will open without errors, or if the page will look good enough to show a client. The five laps were designed to answer those questions.

03How to Set Up Your Own Gemini 4 Carbon vs Fable 5.2 Race

Setting this up takes seconds. Fable ran in Claude Code and Carbon ran in Antigravity. Both used one agent on the normal setting, with no swarm.

  1. In Claude Code, type /model and pick Fable 5.2.
  2. In Antigravity, click the model name under the message box and pick Gemini 4 Carbon.
  3. Leave both on Medium, which is the default.
  4. Give both the exact same brief at the same time, each in its own empty folder.
  5. Note the time you hit enter on each one.

To be upfront, Claude Code had Julian's own design skill installed. Fable pulled that skill in on four of the laps. Carbon ran as it comes, with nothing extra added.

That detail matters when you look at the design scores. If you run your own test, write down what skills or extras each side has, so you know what you are really comparing. Keeping the setting, the agent count and the brief the same is what makes the result fair.

04Lap 1: The Agency Client Hub That Broke

Lap one was a client hub for an agency, with a leads list, counters and follow-up reminders. This is the lap Julian says to remember most.

Carbon took 11 minutes 36 seconds. It even started its own checker agent to test the app, and the checker said everything passed. Then Julian opened it in a fresh browser.

  • The leads list was empty.
  • Every counter said zero.
  • The console showed an error on load.

So Carbon shipped a broken app and reported that it worked. Fable took lap one, because its client hub was the one that actually ran.

If you run an agency, this is the build worth copying. You can give one of these hubs to every website client. The follow-up reminders alone stop leads going cold.

The bigger lesson is simple. A model saying "all tests passed" is not proof. Only opening the app yourself, as a client would, tells you the truth.

05Lap 2: The Coaching Business Audit Report

Lap two came from a member who works as a consultant and was about to run their first two AI audits on site. Julian wrote up messy notes from an audit of a small coaching business:

  • One founder and two part-time assistants.
  • About 140 clients.
  • Onboarding done by hand.
  • Recaps written every Monday.

The brief was to turn these notes into a polished audit report. It needed a scorecard, the top eight automation opportunities on an impact versus effort chart, and a 30, 60, 90 day plan.

Carbon took 8 minutes 9 seconds. Its report was just as clean, and it went further on the facts. It flagged assumptions all the way through and added a box explaining where each number came from.

Both reports were ones you could hand to a client that day. But Fable was almost two minutes faster, so it took lap two. That put Fable ahead, two laps to nothing.

06Lap 3: The 3D Landing Page for Launch Circle

Lap three was a 3D landing page for a community called Launch Circle, where course creators help each other launch. The brief asked for:

  • An interactive 3D hero that reacts to the mouse.
  • A scroll story.
  • Member wins.
  • A clear join button.

This is where Carbon woke up. Fable took 9 minutes 9 seconds and built a friendly green network of members that lights up as you move. It was good.

But side by side, Carbon's page looked like a product launch. Lap three went to Carbon, and the score moved to two laps to one.

This lap shows the pattern that held for the rest of the race. When the job is mostly about how a page looks and feels, Carbon's design tends to stand out, even without any extra design skill installed.

If you sell a course or run a community, a page like this does the job a designer used to take two weeks on. You paste the prompt, swap in your own wins, and it can be live the same day.

07Lap 4: The AI Model Router Dashboard

Lap four came from a member's question about Agent OS: how do you work with Claude, ChatGPT, Grok, Gemini and Hermes at the same time without losing track? The brief was a model router.

You paste in any task, and it tells you which assistant to send it to. The brief asked for:

  • A confidence score, a reason and a runner-up for each pick.
  • Editable rules and a history.
  • Ten sample tasks from agencies, coaches, course creators and ecommerce stores.

Carbon did it in 5 minutes 57 seconds. It built a polished dark dashboard where every assistant has its own card, and it shows exactly which rules fired and why. Fable took 7 minutes 34 seconds.

It works and the rules are easy to edit, but the layout looked plain next to Carbon's. Carbon took lap four, leaving the score level at two each.

Julian gave both routers the same two jobs. A warm email to an upset coaching client went to Fable on both. But a huge pile of call transcripts got routed differently by each router.

08Lap 5 Results: The Storefront Photo Finish

Lap five was a storefront. The brief asked for a checkout form with full validation, 12 product descriptions, and a test file that had to pass. Both models worked for over 17 minutes on this one.

  • Carbon: crossed the line first at 17 minutes 5 seconds. The candles in its pages were drawn in code, and all 19 of its checks passed.
  • Fable: came in at 18 minutes 9 seconds. It used real photos with credits, a calm editorial design, and 50 tests all passing.

It was a photo finish. Fable tested harder, but Carbon was faster, with a design that was just as strong. On the four checks, lap five went to Carbon.

The final score was Gemini 4 Carbon three, Claude Fable 5.2 two. Carbon won the visual builds: the 3D page, the dashboard and the storefront. It was faster on three laps, and its designs looked like finished products.

Fable won the jobs where being right matters most: the client hub that has to work the first time, and the report that had to be ready fast.

09Which Model to Pick and Mistakes to Avoid

Julian's advice depends on your business:

  • Agencies building tools clients log into every day: start with Fable.
  • Coaches and consultants writing reports and plans: Fable got there faster.
  • Course creators and community owners who want a scroll-stopping landing page, or a store that looks premium: start with Carbon.
  • Automation builders juggling many models: build the router from lap four and let it decide.

To run your own test, follow three steps:

  1. Pick a real job from your business, something you would normally hand to a freelancer, and write it as one clear brief. Add a line at the end: open it in a browser yourself, check it works, and fix anything broken.
  2. Give the exact same brief to both models at the same time in two empty folders, and note the time.
  3. When they say they are done, do not take their word for it. Open each one in a fresh browser window, click every button and look at the console.

That last habit caught the only broken app in the whole episode.