Canva switched on Canva Craft, its first in-house image model, for every user on October 22, surfacing it inside the Dream Lab panel without a paywall or waitlist. The model was built from scratch by Canva Research, the 160-plus-scientist team that absorbed Vienna’s Kaleido (acquired 2021) and Leonardo.ai (acquired 2024, whose researchers TechCrunch reports contributed to Craft). It sits inside a 400-person AI unit now shipping against OpenAI head-on.

The structural choice is the story. Craft was trained on millions of human-designed templates rather than photographs, and it runs a two-stage pipeline: a planning model interprets the prompt for layout, hierarchy, and typography before a rendering model produces up to 2K output across 30-plus aspect ratios in 55 languages. That’s the opposite of how general-purpose image models have been built, and it addresses the one failure mode small-business owners have never been able to work around: broken text and awkward composition on a flyer, menu, or Instagram post.

Early benchmarks suggest the bet landed. On DesignArena’s crowdsourced leaderboard, Craft debuted fourth overall behind only OpenAI’s GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2, with an Elo of 1340 at launch that settled to 1330 by October 23. In category play it ranked first among 87 models in People & Portraits, first among 85 in Landscapes & Nature, second in Abstract & Patterns, fourth in Products. Canva says its designers assessed over 25,000 image pairs during fine-tuning. “When automated evaluations and human judgments differed, we sided with the designer,” the company wrote.

A Craft job spends user AI credits and takes roughly 50 to 60 seconds end to end, per Canva’s developer forum. Faster variants, Craft-fast and Craft-ultrafast, are promised “at a fraction of the cost of leading frontier models,” which matters at Canva’s scale of over 400 designs per second. Stef Corazza, Canva’s head of AI research, put the economics plainly: “even a variation of a few cents per generation makes an enormous difference.” A Canva Connect API preview (craft_v1_2k) and an MCP Server for ChatGPT and Gemini are listed as coming.

For a founder, the practical shift is that Magic Layers converts a generated poster or social post into an editable Canva design rather than a flat file. That compresses the gap between idea and finished asset, and it slots into the same owner-operator stack reshaped recently by HeyGen’s production-free video pipeline and by ChatGPT becoming a discovery surface for small businesses.

LemonLime is on the deciding side: a done-for-you service that studies a small business, its competitors, and its content, then prepares finished branded content and outreach proactively for the owner to approve. Craft is on the production side; the two don’t overlap. What Craft changes is what a non-designer can actually produce once the brief exists.

Sources