Karyakarya

Guides

Claude Fable 5.2 Shipped This Week - Using Its Longer Context to Review a Whole Storyboard at Once

Edison Chua5 min read
Guides

Anthropic released Claude Fable 5.2 on November 9, 2026, alongside Claude Mythos 5.2 - a more restricted sibling aimed at vetted organizations working in sensitive technical areas. The two are reportedly the same underlying model; Mythos simply ships without certain safeguards and a different access process.

The headline specs for Fable 5.2: a 1 million token context window, support for image input alongside text, and output speed that several outlets report as noticeably faster than Fable 5.1. None of that is a video or image generation feature. But for anyone building AI video shot by shot, the context window and image input together solve a specific, annoying problem.

The problem: multi-shot prompts drift

A 20-second AI video is rarely one generation. It is usually several shots stitched together - an opening shot, a product shot, a close-up, a closing shot - each generated separately because every video model still caps clip length well under a minute. Keeping the same character, outfit, lighting and setting consistent across all of those separate prompts is most of the actual work.

With a smaller context window, that consistency has to be managed by hand: you paste the character description into every new prompt request, and by shot 15 something has usually drifted - the chat forgot a detail from shot 3, or you changed one line and forgot to update the rest. A 1 million token window can hold the entire brief - brand notes, reference description, and every shot written so far - in one conversation without needing to re-paste it each time.

Where image input adds a second layer

  • Paste several reference photos of a product, character or location at once and ask for one consistent description to reuse across every shot prompt, instead of writing that description from memory and having it drift shot to shot.
  • Paste a batch of frames you already generated and ask it to flag inconsistencies - a changed outfit, a different lighting direction, a camera angle that breaks continuity - before you spend credits generating the next shot in the sequence.
  • Hand it a messy shot list plus a handful of mood-board images and ask for a cleaned-up, numbered prompt for each shot, written in the level of detail a model like Seedance or Kling actually needs (shot type, camera move, lighting, what happens and in what order).

What it does not do

Fable 5.2 is a text and reasoning model with image understanding, not a media generator. It will not render your video or your image - it will help you plan, write and check the prompts you then hand to a model that does. Anthropic's own framing for this release leans toward coding and agentic work, not content creation, so most of what is useful here is a side effect of the longer context window and image input, not a feature built for this use case specifically.

Pairing it with Karya

The workflow splits cleanly: draft and review your full shot-by-shot script and prompts in a model like Fable 5.2 - including checking reference stills for consistency before you commit - then bring each finished prompt into Karya to generate the actual clip or image. Karya's Video and Image studios run Seedance 2.0, Kling 3.0, Nano Banana Pro and Gemini Omni among others, with the credit cost shown before each generation, so the planning and consistency-checking stays cheap and the credits go toward shots you have already reviewed.

Try it in Karya

Generate images and videos with frontier AI models from one canvas. Subscribe from $19/mo, or top up from $5 - 200 bonus credits with your first purchase, and top-up credits never expire.

Start free