Architecture

The design decisions behind Fable UI, and how it compares with GPT-6 Intelligent UI.

The pieces

components/fable-ui/
  schema.ts        Zod catalog: every block, layout and action, with bounds
  prompt.ts        System prompt that teaches composition
  tool.ts          The render_ui AI SDK tool
  safety.ts        URL, image, filename and MIME policies
  resolve.ts       Turns a partial spec into renderable blocks
  to-markdown.ts   Deterministic Markdown fallback
  provider.tsx     FableProvider: actions, image policy, widget state
  renderer.tsx     FableRenderer
  tool-part.tsx    FableToolPart for useChat message parts
  blocks/*.tsx     The components
  chat/*.tsx       Optional chat shell

schema, prompt, tool, safety, format, resolve and to-markdown import no React and run on the server.

Decisions

One tool, many blocks. The model calls a single render_ui tool with a list of blocks, instead of one tool per component. It composes freely, and you register one tool.

Nested, two levels. A spec is a list of blocks; layout blocks (columns, section, tabs) contain content blocks and never nest. Nested JSON streams in reading order, so a block can render the moment it closes. Two levels keep the JSON Schema non-recursive, which every provider accepts, and keep layouts readable.

A compact tool schema. The schema sent to the model describes each block once; layout children refer to the top-level shapes. Fable still validates every child strictly before it renders. This keeps the tool definition about four times smaller.

Lenient shape, strict bounds. Unknown keys are stripped rather than rejected, so a slightly-off model still renders. Every field has a maximum, and each block is validated on its own.

Container queries. Blocks respond to their own width, not the viewport, so the same chart works full-width and inside a column.

State that survives streaming. Blocks are keyed by position, and widget state (sorting, checklist ticks, form answers, edited drafts) lives in FableProvider, scoped to the tool call.

Compared with GPT-6 Intelligent UI

ChatGPT's Intelligent UI (October 2026) answers with native components from a fixed kit, compiled progressively as the model streams. Public reporting and teardowns describe a model-written mix of Markdown, JSX-like tags and JavaScript, compiled on the server and run in a sandboxed iframe and worker, with a Markdown fallback stored alongside. Weak spots called out by reviewers: layout shift while streaming, no navigation in long answers, an invisible capability ceiling, and unclear state persistence.

GPT-6 Intelligent UIFable UI
Where it runsInside ChatGPTYour Next.js app, with your model provider
ComponentsClosed kitSource code in your repo; add, remove or restyle blocks
LookChatGPT'sYour shadcn tokens, or six presets
Model outputMarkup plus JavaScript, compiled to codeJSON, validated; nothing executes
SandboxRequiredNot needed
Layout shiftReportedSkeletons hold each block's space; media never renders from a partial URL
Long answersNo navigationArticle contents, tabs and FAQs
Back to the modelNew user turnreply actions and form submits, as visible messages
Side effects—Only actions you register, with optional confirmation
FallbackStored MarkdownspecToMarkdown

What Fable deliberately gives up: generated mini-apps with arbitrary logic (ChatGPT's AppBlock). Fable's interactive blocks (forms, checklists, templates, editable drafts, sortable tables) cover the common cases without executing model code.

How a given model performs with Fable depends on the model; measure it with the evaluation.