OpenAI launches Canvas, a multimodal model tuned for design workflows

AI · 5 min read

OpenAI says Canvas blends image, vector and structured UI understanding into a single model that designers can query conversationally. The company demonstrated workflows where a designer uploads a rough hand sketch and Canvas returns a layered Figma file, CSS snippets and a list of accessibility improvements.

Canvas also integrates a new “handoff mode” that generates component specs, token mappings, and platform-specific code snippets (React, SwiftUI, Jetpack Compose) alongside annotations for interaction states. OpenAI emphasized on-device inference paths for privacy-sensitive projects and a suite of plugins for major design tools.

The release focuses on practical constraints: Canvas limits hallucination by returning verifiable assets and citations to design tokens or external libraries it used. OpenAI is shipping Canvas as both an API and a Figma/Adobe plugin in an invite-only beta, with wider availability promised later in the year.