Yana Lialuke
Claude Fable 5.2 came out on Nov 9. We gave it a brief Fable 5.1 had already built: a kid's bedroom at night with a hidden monster, all in three.js code. Fable 5.2 built the whole scene in one pass, in 1h 32m and about $27.50 at API prices. Fable 5.1's first pass took 1h 14m and about $28.03. In a second test, Fable 5.2 rebuilt the Shrek opening in voxels against GPT-6.1 Sol: about 7x faster, but 4.7x more expensive.
The setup
Each model got the same brief and worked as a background agent on a Claude subscription. It wrote the code, took screenshots of its own scene in a headless browser and fixed what it saw.
- Models: Claude Fable 5.2 and Claude Fable 5.1
- One live three.js scene per model, with every prop, texture and light made in code
- A 48-second walk through the room with a flashlight, like a kid looking for a monster
- References: photos of cluttered kids' rooms and frames of moonlit bedrooms from animated films, for the look only
- Test dates: Nov 9–10, 2026
Both ran on our subscription, so the dollar figures are what the same tokens cost at the API list price. Fable 5.2 is also cheaper per token: $8 per million input and $40 per million output, against $10 and $50 for Fable 5.1. More on the new model is in Claude Fable 5.2 and Mythos 5.2: benchmarks, price and safety.
The brief
The room had to look like a set from a big-studio animated film: soft, warm, rounded and lived-in. The mood is cozy horror for kids, a little spooky and never gory.
- A fully closed room of about 4.5 by 4 meters, with a bed, a wardrobe, a desk with a retro computer, shelves, a window and a rug
- At least 150 props and no empty walls, shelves or corners
- Moonlight through the window as the main light, plus flickering lamps, screens and fairy lights
- A crayon drawing of the monster on the desk
- One fake scare, then the monster's reveal near the end
Each model also got its own monster brief and never saw the other's:
- Fable 5.1 – a furry, horned creature in warm colors, hiding in the wardrobe
- Fable 5.2 – a scaly creature in cold colors, hiding under the bed
The numbers
The first full pass is the fairest match: one agent, one brief, no extra rounds.
| Fable 5.2 | Fable 5.1 | |
|---|---|---|
| Time, first full pass | 1h 32m | 1h 14m |
| Cost, first full pass | $27.50 | $28.03 |
| Output tokens, first full pass | 365,528 | 171,964 |
| Code in the final scene | 2,705 lines, 21 modules | 3,877 lines, 14 modules |
Two rooms
Both rooms are full: posters, toys, books, a quilt and glow-in-the-dark stars. Fable 5.2's room is brighter and more colorful, with patterned wallpaper, bunting and a pink desk. Fable 5.1's room is darker and more blue, with a heavier moonlit look.
Both models chose the same fake scare: a coat on a stand next to the wardrobe. Fable 5.2 added a hat and glasses that catch the flashlight, so it looks like a figure for a second.
Two monsters
- Fable 5.2 – named its monster Squelch: a teal salamander with violet bands, magenta gills and amber eyes. It shows eyes and claws in the gap under the bed, then slides out, stands up and grins. In its own notes, Squelch is afraid of the vacuum cleaner and only eats the left sock of a pair
- Fable 5.1 – built Grimble: a furry, horned orange creature with many eyes that leans out of the wardrobe and lunges at the camera
Squelch's pose has one weak spot: an arm stretched out on the rug looks odd. Overall it reads more like a funny lizard than a scary monster.
Fable 5.2's scene also runs faster: 236 frames per second on our machine, against 165 for Fable 5.1.
A second test: Fable 5.2 vs GPT-6.1 Sol
On Nov 8–9, we ran Fable 5.2 against GPT-6.1 Sol on another scene. The task was the 21.5-second opening of the "All Star" sequence from Shrek, rebuilt in voxels. Both models got the same reference video and wrote their own code. Fable 5.2 ran at Extra effort and Sol at xhigh, both on subscriptions. We posted the side-by-side video on X.
| Fable 5.2 | GPT-6.1 Sol | |
|---|---|---|
| Time to build | 29m | 3h 26m of agent work |
| Output tokens | 712,456 | 454,801 |
| All tokens processed | 44.8M | 44.7M |
| API-equivalent cost | $56.89 | $12.02 |
Both models did about the same amount of work: 44.8 million against 44.7 million tokens, and 94–95% of them were cache reads. Fable 5.2 wrote 1.57x more output. More than half of Sol's output, 244,880 tokens, was reasoning.
The 4.7x cost gap comes mostly from price, not volume. Fable 5.2's list prices are 4x Sol's on input and output. Output was half of Fable's bill at $28.50, and cache writes added another $19.97.
Fable 5.2 finished in about 29 minutes, against about 3.5 hours for Sol. We did not score the two scenes, and the video is the comparison. Fable's version has two known flaws: the SHREK title runs to the edges of the frame, and two shots show a dark strip where the set ends.
Our take
Fable 5.2 built a full, dense room with its own monster in one pass, at a lower price per token, and its scene runs faster in the browser. Against GPT-6.1 Sol, it was about 7x faster but 4.7x more expensive: Fable 5.2 saves time, not money.
Sources
- Claude Fable 5.2 and Mythos 5.2: benchmarks, price and safety – our news on the release
- three.js – the 3D library both scenes use
- Brief, runs, timings and token counts: our test, Nov 9–10, 2026
