Pure Code Generates a Ghibli Animation—and Also Reveals the Unreleased Fable 5.5

Chetaslua posted an animation on X. Sunrise, an abandoned forest, a moss-covered giant robot rising from the ground, birds startled into flight, engines igniting, soaring toward a golden sky. Ghibli-style cel shading, hand-drawn lines, watercolor-textured backgrounds.

He wasn’t kidding. It really was generated by a single HTML file.

Baoyu reposted the tweet with a restrained assessment: Fable 5.5’s video capabilities have taken another leap; getting pure code to this level isn’t easy, but the benefit for ordinary users is limited.

Break that sentence down: the first half is fact, the second half is judgment.

One file, no external assets

Chetaslua says the whole thing uses no external resources. The final product is a self-contained index.html with no dependencies and no build step; download it, open it offline, and it plays. The MP4 is embedded in the page as base64 and decoded into a Blob on load. The player is custom-built, with play/pause, a progress bar, fullscreen, and a sound toggle; if autoplay fails, it falls back to muted. There’s also a <canvas> layer that generates film grain in real time. Keyboard controls: Space/K play/pause, ←/→ back/forward 5 seconds, M mute, F fullscreen.

The code is open source on GitHub, and there’s an online version you can view directly:

• Repository: https://github.com/ChetasLua/moss-keeper
• Watch online: https://chetaslua.github.io/moss-keeper/

The prompt is extremely detailed. It asks for a single self-contained HTML file as the final deliverable, and allows any tools, libraries, or references. The art style is locked to Ghibli; the protagonist is an original giant robot; no people appear in the scene. Then comes a long list of “must not look like” rules: no 3D plastic materials or CG lighting; no SVG-style perfect curves or uniform strokes; no web-animation-style sliding layers or software easing; no shape distortion or melting details common in AI art; no Western cartoon style; no stiff poses or empty backgrounds.

The last line is the acceptance criterion: before finishing, randomly sample frames and inspect them as stills; if any frame doesn’t look like a theatrical animation, keep revising until every frame could pass as a screenshot from a feature film.

The result is the clip above.

How the new model was detected

Fable 5.5 hasn’t been officially released yet; the author likely got access through a staged rollout. His verification method is a bit ruthless: turn off web search and memory, then ask the model directly:

Where did @karpathy go in 2026?

Only the new model knows the answer. If it knows, its training cutoff is recent enough; if not, it’s an old model. Essentially, it’s probing the knowledge cutoff date.

In the screenshot, the model answers: Andrej Karpathy joined Anthropic in May 2026 to do pretraining research and use Claude to accelerate the research process. The answer is followed by a note that this is from memory, unverified, and details may be inaccurate.

The author added a reminder: a knowledge cutoff date doesn’t mean the model knows everything before that date. Some events happen before the cutoff and the model still doesn’t know them. This test can conveniently reveal that too.

Give the same prompt to the regular Opus 5.5, and quality drops. On Fable 5.1, it’s even worse.

Two reality checks in the comments

One comment says the video is still fundamentally pixels; for this generation method to catch up with the quality of video models, the compute required would have to increase by who knows how many times. Another is more practical: what ordinary users really lack is local editing—changing one spot without rerunning the whole video. Only when that’s solved does it count as a real boost.

The second point hits the nail on the head. Looking at these demos now, impressive as they are, the most expensive part of the workflow has never been generating from scratch—it’s editing. If the lighting in one shot is wrong or the timing of one action is off, you rerun the whole segment, and the time cost immediately eats back the “efficiency gain.”

Baoyu says he’s more looking forward to improvements in other capabilities, which is probably the same point. Maxing out video capability is one thing; whether ordinary people can actually use it is another.

Getting animation to this level with pure code is definitely worth a look. But what determines whether it stays in your daily toolchain is how long it takes the next time you only want to change three seconds.