21+ prompts. 145 things people made with Claude Fable 5.2 in its first week, from Blender films and Three.js scenes to games, fight animations, music and paintings, each credited to its creator and linked to the original post, with the prompt where the creator published one. Fable 5.2 does not draw pixels: every piece here was rendered from code it wrote.
← All 145 Claude Fable 5.2 prompts
Last updated: · Claude Fable 5.2 API — pricing & playground · API docs
Claude Fable 5.2 (model id claude-fable-5-2, released by Anthropic on 9 November 2026) is a language model: it takes text and images and returns text, with a 1M-token context window and up to 128K output tokens. It does not render pixels or synthesise sound by itself. Every film, game, song and painting on this page came from code it wrote, such as Blender Python, Three.js and raw WebGL, Canvas, Web Audio, numpy or LilyPond, which an agent like Claude Code ran, rendered and usually encoded with FFmpeg. Many of the best-known results started from one or two sentences, but the runs behind them were long: the pelican film wrote about 4,800 lines of Python in a session of about 1.1M output tokens, and the FPS level ran for 8 hours at about $550 by its author's account. The 145 cases were posted on X between 4 and 11 November 2026; the few posted before 9 November are pre-release tests their authors labelled "Fable 5.5". 38 cases have the prompt as the creator published it (9 long briefs are excerpted, with our outline), 20 have only the creator's one-line request, and 87 were never published. On apimodels.app, Fable 5.2 costs $4.80 per 1M input tokens and $24 per 1M output tokens, against Anthropic's $8 / $40 list.
Source: Anthropic · Claude Fable 5.2 (the model vendor’s official site)
Only the creator's request is public, not a full prompt: "create a painting as indistinguishable as possible from a photograph"
No prompt published. Before painting, the model first gives itself a painter persona, and each painting runs to 1,000–2,000 lines of code.
Only the creator's request is public, not a full prompt: "make a masterpiece"
Prompt: I want you to create or find a tool you can use to generate an oil painting that allows you to take a video of you making each stroke. Then I want you to make a painting and video it of what it feels like to be you.
Create the most complex original painting you can. Like on the greatness of Leonardo Davinci
No prompt published. It wrote its own brush engine and invented a painter, Odile, a groundskeeper at a tennis club in 1978, then painted 6,300 strokes in Python.
No prompt published. The life of a single seed is used to tell a story about karma: 13 paintings, roughly 210,000 strokes, 74 seconds.
No prompt published. An oil painting of a Swiss lake scene that took 1 hour 11 minutes and cost $22.62.
No prompt published. A painting engine built fully autonomously: 2,101 lines of TypeScript supporting oil, watercolor, pencil, charcoal and pastel, used to make the piece 《有人醒着》 ("Someone Is Awake"). It is open source at github.com/SweetRetry/atelier.
No prompt published. The Mona Lisa recreated in pure JS code (see also the 62,430-stroke version by @AnikaCodes, post 2120467261458911412).
No prompt published. A portrait of Sam Altman painted on a digital canvas that simulates how real oil paint mixes, working from a reference image and taking screenshots to check its progress as it went.
No prompt published. A piece in which every single brushstroke was made by Fable 5.2.
Only the creator's request is public, not a full prompt: "animate its own life, from day 0 to now"
No prompt published. Pixel art in an Impressionist style (60 seconds).
No prompt published. A photo of Bora Bora repainted with 311,905 strokes, ending up almost identical to the original. The post relays someone else's work rather than the poster's own.
Prompt formula
Creative coding brief = the deliverable and its format (an MP4 at a set size and frame rate, or one self-contained page) + the design pinned in numbers (sizes, frame counts, tempo, budgets) + a build contract (renderAt(t) or a per-frame state, which renderer) + self-checks on stills before the final render + stops and a cost gate before anything paid + permission to take the time it needs
Example (our template, not a creator's prompt): Make a 20-second Blender shot at 1280x720 and 24 fps: a lighthouse on a rock at dusk, waves breaking, the beam sweeping through the spray, a slow push-in. Build everything in Python from an empty scene and compute every keyframe from the frame number. Render 4 preview frames at 25% size, look at them and list what is wrong before each change. Before the final render, tell me the seconds per frame and the total time, and wait for my OK. Deliver one H.264 MP4. Take the time it needs.
Not directly. Fable 5.2 takes text and image input and returns text. Every piece on this page was rendered from code it wrote: Blender Python rendered by Blender, Three.js or WebGL pages captured by a headless browser, Web Audio or numpy for sound, canvas stroke engines for paintings, and FFmpeg to encode the result. In a chat window you get code or an HTML page you can open; to get an MP4 you need an agent that can run commands, such as Claude Code.
Some did. The FPS level, the pelican film and the browser Minecraft clone started from one or two sentences. But one prompt rarely means one quick reply: reported run times range from about 20 minutes to 8 hours, one open-world game was built over 7 days, and several creators added follow-up messages. Other cases ran on briefs of about 6,000 to 9,400 characters. Every card says whether the text is the creator's prompt, an excerpt of it, only their one-line request, or unpublished.
As the creators reported them: $7.48 and $8.43 for the two pixel-art switches, $21 for the Stonehenge scene, $22.62 for a Swiss lake painting, $34.96 for a 2,000-line Three.js short, and about $550 for the 8-hour FPS level. A 4K Blender render on 10 rented GPUs cost $5.75 in GPU time. This kind of work is output-heavy: the pelican run wrote about 1.1M output tokens, which is about $26 of output on apimodels.app at $24 per 1M, or about $44 at Anthropic's $40 list price.
An agent that can run commands (Claude Code or similar) pointed at Fable 5.2; Node.js with Playwright and FFmpeg for anything rendered in a browser; Blender 5.x for the Blender cases, plus the Blender MCP add-on for the cases that used it. A GPU helps for Cycles, but EEVEE runs on a laptop: our own starter script renders 72 frames at 1280x720 in 77 seconds on an Apple M4. Music cases use Web Audio in the browser, numpy or LilyPond.
Fable 5.2 defaults to high effort on the API; Claude Code and the Claude apps default to medium. Raise it for long one-shot builds where a failed run costs more than the extra tokens, and keep medium for small fixes and iterations. Creators on this page used medium, High and Extra High.
Set ANTHROPIC_BASE_URL=https://api.apimodels.app, your apimodels.app key in ANTHROPIC_AUTH_TOKEN and ANTHROPIC_MODEL=claude-fable-5-2, and Claude Code runs on Fable 5.2; the same key works on /v1/messages and the OpenAI-compatible /v1/chat/completions. If the primary upstream errors, requests switch to a second channel automatically, and failed requests are not billed. It is billed per token: $4.80 per 1M input, $24 per 1M output and $0.12 per 1M cache reads, with no subscription.
The works belong to their creators. Clips from X play straight from X's own servers, every card links to the original post, and we do not host or re-upload any of them. Published prompts are quoted with credit and a link: in full when short, excerpted when long, and never filled in when unpublished. Check the original post before using a prompt commercially. If you made one of these and want the card removed or the credit corrected, email us.
Claude Fable 5.2 is live here alongside 160+ other image, video, audio and language models — one API, one key, pay as you go, with $0.10 of trial credit on sign-up.
The whole collection is on GitHub as Markdown and JSON: all 145 cases with credits and links to the original posts, our outlines of the long briefs, and two starter scripts we ran ourselves (Blender and Three.js). The videos and images stay with their creators and are linked, not re-hosted. Prompts up to 400 characters are quoted in full; longer ones are excerpted, and unpublished ones are left out. Our own write-ups, outlines and code are licensed CC BY 4.0. Open an issue to add a case or correct a credit.
View on GitHub