Blog · 2026-11-11

Claude Fable 5.2 Use Cases: 8 Ways People Make 3D, Games, Music and Paintings With Code Alone

Model page: Claude Fable 5.2 API — pricing, specs and the playground

What can Claude Fable 5.2 do? Beyond coding and analysis, its most talked-about trick is producing finished work by writing code: Python that drives Blender to render a 3D animation, a home-made paint engine that lays down a 700,000-stroke photoreal painting, a playable shooter built in eight hours, all without calling an image or video model. The cost is time and tokens: big pieces run for hours and cost tens to hundreds of dollars. Here it is $4.80 / $24 per 1M tokens, 40% below Anthropic's list price.

Claude Fable 5.2 is the model Anthropic released on November 9, 2026, replacing Fable 5.1. It has a 1M-token context window, up to 128K output tokens per call, text and image input, and thinking that is always on, with an effort setting for how hard it thinks. In the two days after launch, X filled up not with benchmark charts but with work people could hardly believe was made in code. The Chinese tech writer Khazix (卡兹克) summed up the reaction: the surprise is not that it writes code, but that what the code makes has taste.

This guide sorts the work into eight kinds of use. For each one we cover what was made, how, how long it ran, roughly what it would cost at our prices, and what to watch for if you try it. The examples are pieces developers shared publicly on X. We credit each author, checked the numbers against the original posts, and do not reproduce their images or videos.

A word on who we are: we run apimodels.app, a platform where one API key calls Claude Fable 5.2 alongside Opus, GPT, Gemini and other language models, plus image, video and audio models. The work described here is not ours; the costs are worked out from the usage the developers published and our prices on November 11, 2026.

The short version

What these uses share is that Fable 5.2 turns pictures, motion and sound into code, then runs that code, looks at what it rendered and fixes it. An image model gives you one picture per roll and a fresh roll for every change. Code can move one feather, one camera cut or one note precisely, and run again as often as needed. That makes it strongest where you need control, consistency and endless revisions: 3D scenes, web animation, games, vector art and procedural painting.

The barriers are just as clear. Most of this work ran inside an agent setup such as Claude Code, for one to several hours per piece, and Blender projects need Blender installed locally, ideally with a GPU for rendering. If you only need one image or a realistic video clip, an image or video model is faster and costs far less.

Why code can make paintings, animation and music

Inside a computer, an image is an array of pixels, a 3D scene is a pile of coordinates, materials and lights, and music is a sequence of pitches and durations. All of it can be generated by a program. Earlier models could write such programs, but the results looked clumsy. What changed with Fable 5.2 is taste: it knows how feathers should lie, where the light should come from and where to cut the shot. Combine that with long working sessions and the habit of inspecting its own renders, and one person with one model can run a pipeline that used to need a modeller, an animator, a lighter and an editor.

1. Driving Blender with Python to make 3D animation

The best-known example is Tech Friend AJ's pelican riding a bicycle. His request was one paragraph: make a video of a 3D pelican on a bicycle, as realistic as possible, with believable anatomy and feathers, a real drivetrain that moves, and real-world lighting and camera work. Inside Claude Code, Fable 5.2 wrote about 4,800 lines of Python that drove Blender from the command line with no interface open. It generated the mesh and feathers in code, rigged the bird with 21 bones and computed the skin weights, built a 14-inch child's bike whose chain turns at 66 rpm, and rendered 444 frames on a GPU. The session ran about 2 hours 44 minutes and produced about 1.1 million output tokens; he published the full trace on Hugging Face.

There are plenty more. Jeremy Berman posted an 18-second 3D short made the same way, with Fable writing Python to control Blender. 0kncn built a 90-second, 42-shot 2D anime piece by assembling scenes through Blender's Python API, rendering in Eevee and tracing line art frame by frame. Stefan_3D_AI asked for a sea serpent coiling around a ship for 20 seconds without ever passing through the hull.

If you try it, three habits help: run Blender headless from the command line; render a small test first and have the model report the time per frame before the full render; and tell it to look at a render after each step and fix any clipping or wrong proportions itself. At our prices, a job the size of the pelican costs about $26 in output tokens alone, while most of the input hits the cache at $0.12 per 1M tokens.

2. Three.js web scenes you can export as video

If you would rather skip Blender, ask for a single self-contained Three.js page that opens as a 3D scene in any browser. RezAIT00 asked for a floating voxel island of Paris in 1888 with the Eiffel Tower under construction, 14 seconds long and already moving in the first frame. His prompt included one very practical requirement: the page exposes a renderAt(t) function, so every frame can be captured at an exact time and stitched into a video.

The web route goes deep too. Deedy Das asked, in a single sentence, for the most intricate photorealistic honey simulation it could write in WebGL. mylifcc built a wooden-ball maze in about 2,000 lines of Three.js, and the model profiled its own code, cutting frame time from 55 ms to 9 ms. Web pieces cost less, iterate faster and can be shared as a link the moment they work.

3. Painting stroke by stroke in code

The work that stopped people scrolling came from Jake Eaton. He had Fable 5.2 paint without any image model: each painting is a Python program of one to two thousand lines with a simulated paint engine and different bristles, and every stroke passes through that engine before it lands on the canvas. One photoreal piece took about 700,000 strokes, and zooming in shows the brushwork. His method was to have the model invent a painter first, with a name, a backstory and a style, then paint as that painter, revising the brushes and paint engine after his feedback.

It goes further still. BlendiByl made a browser game with a hand-painted look at about 1.5 million strokes per frame, running at 200 fps, built in under an hour and released as open source. To try it yourself, start with the same two steps, invent a painter and then write the paint engine, and state plainly that no image model and no tracing are allowed.

4. Building a playable game from scratch

Felix Rieseberg asked Fable 5.2 for a first-person shooter with graphics that could keep up with AAA games: small in scope, but visually impressive. It worked for eight hours straight with no feedback in between, generating every model, sound and texture in code, and produced Saltline, a browser game anyone can play from a link. He reported about 2.7 million output tokens, roughly $550 at Anthropic's list price; the same usage would cost about $330 here.

Smaller projects are worth a look too. joshvlc asked for a 3D basketball game in Three.js with dribbling, shooting, dunks, steals and passes. AmirMushich built Worm Rider, a 220 KB game of 3,400 lines, in ten prompts and open-sourced it. When you build a game, spell out the rules one by one and ask the model to play it and check screenshots itself.

5. Composing music and sound effects in code

Music can be code as well. One user had it write a Bach-style four-part chorale in LilyPond, the music engraving language, in G minor and 3/4 time, returning code that typesets as a score; another had it write a symphony. wshuyi went further: a 30-second animation of a sabre-toothed cat hunting, with no stock assets and every sound effect generated in code. For now its music reads more like a composition exercise, and a song you would release still belongs to a music model such as Suno v5. For an original score under an animation or a handful of sound effects, it is already enough.

6. SVG animation and motion graphics

Vector art is home turf for code. marcthecreatorr asked for a pelican riding a bicycle through five animated seasons and got back a single SVG file. Khazix set a fuller brief: a 40-to-60-second motion video celebrating ChatGPT's fourth birthday, no voice-over, original music, with every fact researched online and nothing invented. The model worked for close to an hour, handling the research, layout, music and editing itself. Leviwoyun asked only for a 30-second motion piece to serve as a motion designer's portfolio and had a finished film 24 minutes later.

7. Web and interface design

Design comps come out as working web pages. MiaAI_lab made 76 single-file HTML pieces with Fable 5.2, from SVG illustration and Canvas generative art to WebGL shaders, brand landing pages and dashboards, each published with its prompt. What they share is a style nailed down up front, such as a Memphis-style landing page for a hot sauce brand, or a Mars greenhouse console on an amber monochrome 1980s CRT terminal. Tmeister got 20 redesigns of one landing page from a single short prompt.

8. Fable as the director, other models as the crew

Code is good at structure, camera work and timing, and weak at realistic faces and skin, so some creators combine the two: Fable builds a grey-box animation in code and locks the shots and pacing, then a video model such as Seedance 2.5 renders the real characters. Others plug music and voice MCP servers into Claude Code and let Fable score and narrate the piece itself. On apimodels.app, Fable 5.2, Seedance 2.5 and ElevenLabs v3 share one API key, and with our MCP server installed, Fable can call image generation, video generation and text-to-speech tools directly from Claude Code.

Habits that make it work

• Run it in an agent: almost all of this work happened in Claude Code, where the model can write files, run them and see the results.
• Turn effort up: many of the complex pieces ran at xhigh or even max.
• Make it look at its output: render, screenshot and check after every step, and let it fix what it finds.
• Start small: a small render or a short clip first, then the full run once the timing is known.
• Keep prompts short: as Felix put it, "the age of prompt engineering is basically over". Say what you want and how high the bar is.
• Name the style: concrete targets such as "photoreal enough to pass for a photo" or "Memphis style" work far better than "make it nice".

What it costs and how long it takes

Fable 5.2 is billed by tokens. Here it costs $4.80 per 1M input tokens, $24 per 1M output tokens and $0.12 per 1M cached input tokens, 40% below Anthropic's list price, as of November 11, 2026. Most of the published pieces ran for one to several hours. A 3D job like the pelican produced about 1.1M output tokens, so 1.1 × $24 ≈ $26 in output alone; the eight-hour shooter cost about $550 at list price and would be about $330 here. Thinking is always on and counts as output, so turn effort down for small pieces. Failed requests are not charged.

When not to use it

• You need one picture: for a poster, product shot or avatar, use an image model; GPT Image 2.6 Flare starts at one cent for a 1K image and usually returns in about a minute or less.
• You need realistic people on video: real faces, skin and motion belong to a video model; code is better at structure and camera work.
• Budget and time are both tight: big pieces take hours and tens of dollars, and you still have to judge the result. For everyday coding, start with Claude Opus 5.5 at half Fable's price per token.

How to call it

On apimodels.app the model name is claude-fable-5-2, and one API key calls it through the Anthropic-style /v1/messages or the OpenAI-compatible /v1/chat/completions. To try these uses in Claude Code, point Claude Code at our endpoint and pick claude-fable-5-2; the steps are in the Claude Code guide.

Related models

• Claude Fable 5.2: the subject of this guide, $4.80 / $24 per 1M tokens.
• Claude Opus 5.5: better value for everyday coding, $2.40 / $12 per 1M tokens.
• GPT-6 Astra: OpenAI's flagship, $3.40 / $17 per 1M tokens.
• Gemini 4 Argon: Google's flagship, $1.10 / $5.50 per 1M tokens.
• Seedance 2.5: the video model to pair with Fable for realistic characters, from $0.12 per second.

If you are choosing between Fable 5.2 and Opus 5.5, read Fable 5.2 vs Opus 5.5.

FAQ