Grok 4.7 just mogged Claude Fable and nobody's talking about it
Grok 4.7: $0.44, 1h 23min
Fable 5.2: $15.80, 25min
36x cheaper. Yes, it took 3x longer, but for that price I'll happily wait.
Hide the price tags. Which jet ski is from GTA 6? 👀
- GTA 6 - 13 years and a rumored $2 billion for the whole game
- Claude Fable 5.2 - 5 hours and about $2.50 worth of my weekly Claude Max quota (10%)
Give vibe coders a year and they’ll build GTA 7 on 1% of the budget
Build a single self-contained HTML file: an interactive, photoreal 3D scene of a
gummy bear (and a pear) made of glass or jelly, sitting on a brushed-steel table
in afternoon window light. Match the attached reference for glass quality,
transparency and light. All UI text in English. No external libraries, images,
audio files or network requests; it must work offline by double-click.
RENDERING (this is the core, do not use three.js or meshes)
- WebGL2, one full-screen fragment shader, sphere-traced signed distance fields.
- Glass: real refraction (IOR 1.52), Fresnel reflection, up to 4 internal
bounces, Beer-Lambert absorption by path length (thin edges light, thick core
saturated), a thin clear outer skin.
- Jelly: IOR 1.36, cloudy internal scattering that glows when backlit, weaker
absorption.
- Scene: a room with a window wall on the left; sunlight through the panes with
soft leaf shadows; brushed steel table with anisotropic blurred reflections;
a color-tinted shadow with a bright focused caustic spot behind the object;
dappled light on the far walls.
- Quality: when the camera is still, accumulate frames in a float framebuffer
with sub-pixel jitter and thin-lens depth of field; ACES tone mapping, subtle
bloom, vignette. While moving, render one clean sample per frame. Adapt
render resolution to keep it smooth.
SHAPES
- Gummy bear from smoothly blended ellipsoids and capsules: torso, round belly,
head, two ears, snout with nose, pressed-in eyes, arms held out to the sides
with round paws, short legs with flat soles.
- Pear: two blended spheres, faint vertical ribs, top dimple, clear curved stem.
- A Bear / Pear switch that morphs one into the other by blending the two SDFs.
INTERACTION
- Drag background to orbit, wheel or pinch to zoom, arrow keys too.
- In Jelly mode, grab the figure and pull to stretch it like taffy. Grabbing a
paw, ear or leg stretches only that limb, far, without the shape folding;
grabbing the body stretches the whole part. On release it springs back with
a damped wobble. Tap to poke. The whole body also wobbles and squashes.
- Nine colors as round swatches numbered 1-9; number keys pick them; smooth
color transitions. Glass names (Cobalt, Aquamarine, Emerald, Amber, Ruby,
Rose quartz, Amethyst, Crystal, Smoke) and jelly flavor names.
- "Time outside" slider over 24 hours: sun height, direction and color, sky,
shadow length, exposure; moon and stars at night.
- Buttons: Demo 15s, Sound, Auto-rotate, Reset (view, shape, time).
SOUND (Web Audio, synthesized, soft sine tones only, no noise or harsh waves)
- Glass: mellow crystal note on touch; a quiet note per eighth-turn when rotated.
- Jelly: rounded bloop on touch, low tone following the wobble, boing on release.
- Colors 1-9 are the notes A B C# D E F# G A B.
DEMO 15s
- For exactly 15 seconds the time slider runs from start to end while the
colors play Pachelbel's Canon in D, one color per note, with a soft bass:
opening theme, eighth-note variation, then the fast running passage. The
user can keep rotating, stretching and switching material during it.
ALSO
- Responsive down to phone width, keyboard accessible, respects reduced motion,
clear message if WebGL2 is missing, remembers last shape, material and color.
Before answering, check that the shader compiles, the figure is visible on
first load, and every control listed above works.
Opus 5.5 isn't a bad model… so why did it cost MORE and give me a choppy 3D bear?
Same exact prompt, both on Extra:
Fable 5.2: 15 min, $2.84
Opus 5.5: 23 min, $4.14
fable was faster, cheaper and way smoother
does anyone know why this keeps happening?
Opus 5.5 is cheaper on paper than Fable 5.2, so WHY did I pay more?
same exact task, but Opus took almost 2x longer and was more expensive
Opus: 53 mins, $23.61
Fable: 31 mins, $16.61
did Anthropic quietly nerf the usage limits for Opus 5.5?
BLENDER OR THREE.js? A LAUGHABLE WASTE OF TIME.🤡 I CREATED THIS VIDEO FOR FREE.🙀
Here’s how:
1. The free version of ChatGPT generated reference images and wrote prompts.
2. Then ChatGPT sent them to the free Wan image-to-video demo on Hugging Face.
3. Chat itself added audio and sent me the MP4 file.
No manual editing at all. One prompt, and in 4 minutes and 20 seconds, the video was ready.😎😎
Holy sh*t. Fable 5.2 extra made a dancing jelly bear with its own symphony
52 minutes of work, $9 total, and the music alone is worth it.
Sound on
What song should the bear play next?
I’ve got my own benchmark for Fable 5.2 - a jelly hippo 🦛
16 minutes later, I’m tossing it watermelons. It’s already eaten the first one and its face is still wobbling
What could be more reliable than a jelly benchmark?
Everyone says Fable is overpriced. I tested it.
Same yacht idea, two Claude models:
Fable 5.2: 28 mins, $22.38
Opus 5.5: 43 mins, $18.17
$4 more, but you can see where the money went.
Top one looks better, right?
aaa-looking game environments in the browser with Opus 5.5 + Three.js..
the look comes from a rendering pipeline, not from image gen
tuning it with math instead of gut feeling:
- step 1: opus 5.5: build a pbr pipeline in three.js from papers, talks and unreal’s published techniques
- step 2: three.js: fft ocean, physical atmosphere, global illumination, volumetric clouds
- step 3: three.js: gtao, volumetric fog, auto exposure, local tone mapping
- step 4: references: a dozen frames from real games or film stills, picked by hand for weather, time and composition
- step 5: opus 5.5: pick 7–8 sample areas per reference: sunlit grass, shaded grass, sky near the sun, distant water
- step 6: opus 5.5: measure brightness, chroma and hue in each area, then the same areas in your render
- step 7: opus 5.5: least-squares fit across 100+ points to find the parameters with the lowest error
- step 8: three.js: render, measure, adjust, repeat until the numbers match
don’t ask the ai for a “better-looking” scene. give it physics to render and numbers to hit