Canva ships Craft, its own image model, now on the public API
Canva Research debuts #4 on DesignArena — 1st for portraits and landscapes — with 2K output and text in 55 languages. Plus Cursor CLI's latest agent fixes.
Copy markdown
One async call, paid in AI credits
POST /v1/image-generations with model craft_v1_2k, poll the job, then pull the 2K image — the download URL lives 24h, or set asset_upload to drop it straight into the user's uploads. You need a token with the asset:write scope, and each job spends Canva AI credits rather than a per-image fee.
It landed #4 on DesignArena, behind only OpenAI
Craft is Canva's first in-house foundation image model, trained from scratch on millions of its templates. It debuted 4th overall and 1st for both People & Portraits (of 87 models) and Landscapes & Nature (of 85), making Canva Research the second-ranked image lab on the board.
Text in 55 languages, plus editable Magic Layers
Prompt in any language and get legible rendered text in 55 of them — the usual weak spot for image models. Magic Layers converts a generation into editable design elements, and an MCP server is 'coming soon' to let agents produce design-ready images inline.
Cursor CLI's agent now falls back instead of failing
The v2026.10.22 drop makes a subagent whose pinned model is blocked or Max-only run on the parent's model instead of erroring, with the result noting which model ran. It also reads a subdirectory's AGENTS.md on first edit, and Composer 2.5 subagents now bill at standard speed.