Canva Craft plans the design before generating it: its first in-house model ranks fourth, behind OpenAI

By The Verge · October 22, 2026.

Canva has launched Canva Craft, its first in-house image generation model. According to the company's announcement, it is now available to all users worldwide. It can be found in Dream Lab, under the three-dot menu in the left sidebar.

Craft generates images at up to 2K resolution and in more than 30 aspect ratios. According to The Verge, it competes directly with models from OpenAI, Microsoft and Google. It can be used for social media posts and marketing materials, as well as realistic scenes, digital art and environment images.

The key technical feature is its two-stage architecture. The model first plans the layout and then generates the image. According to Canva, this delivers "greater precision, consistency and control." In other words, it is designed for complex images with legible text and objects placed in exact positions, two areas where conventional generators often fall short.

Canva says it trained the model on "millions of human-made templates" from its library: posts, presentations, infographics, menus. That is a proprietary data advantage that is hard to replicate for anyone without a design platform behind them.

Its showing in external evaluations is more modest. According to Design Arena, a crowdsourced (user-voted) leaderboard of image models, Craft ranks fourth. Ahead of it are GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2, all from OpenAI. Canva also says it includes input and output moderation, privacy and safety controls, and reporting tools.

The roadmap includes two more efficient variants, Craft-fast and Craft-ultrafast, as well as connecting AI agents to Craft through the Model Context Protocol (MCP), an open standard that lets agents use external tools.

Our take: what matters is not whether Craft beats OpenAI, because it doesn't, but that Canva no longer depends on third-party models. Whoever controls the model controls the cost per image, the latency and the integration with the tool. More broadly, design apps are tending to turn generative AI into an in-house component rather than a rented API.

The bet on planning before generating is also a product thesis. A design is not just a pretty picture: it requires hierarchy, correct text and composition. If that planning stage really works, the model will be optimized for a designer's actual work rather than for a beauty contest leaderboard. That remains to be proven. Fourth place in a vote-based benchmark measures preference, not usefulness in a real workflow.

The impact on jobs is real and near-term. Producing marketing materials, menus or infographics of acceptable quality now takes seconds, and that puts pressure on the most routine design work. What gains value is judgment: knowing what to communicate, to whom and with what identity. In the long run, letting anyone or any small business bring a visual idea to life without paying a studio is one more piece of the abundance we advocate: less of a barrier between imagining something and seeing it made.

Two things are worth watching: whether the model maintains its quality outside the templates it was trained on, and what happens when agents start designing without human oversight via MCP.