← Back to blog
AiAbout 6 min read

Canva Built a Model That Doesn't Draw. It Writes Code to Move Your Layers.

Published Nov 14, 2026
Canva Built a Model That Doesn't Draw. It Writes Code to Move Your Layers.

Canva released Craft on October 22, its first image model built in house. Seven days later it released Brush, and Brush does not generate images at all.

That gap is where the story sits. Craft answers the question every design tool has been asked for two years: can you make the picture. Brush answers a different one, the one that shows up on the third revision of a client deck when ninety percent of the layout is right and the headline is wrong.

Brush is a vision-language model trained to edit designs. It reads how a design is assembled, writes small code snippets to carry out the change you asked for, then inspects the result to see whether the change landed. That last step feeds the next one, so the whole thing runs as an agentic loop rather than a single render.

Canva reports the model completes general design editing tasks up to 69 percent faster than its previous design agent, at 46 percent lower cost per task. Those numbers compare Brush against Canva's own prior system, not against GPT Image or Nano Banana. Read them as an internal efficiency claim, which is what they are.

The unit of work stops being the generation

Most image models treat a request as a one-shot: prompt in, pixels out, start over if you don't like it. Designers don't work that way, and neither do the tools they use. A slide has layers. A poster has a type hierarchy. Changing the date on an event flyer should touch the date and nothing else.

A loop of clear film resting under a brass-rimmed magnifier, two frames overlapping.

Canva's launch post makes this the central argument. Creating a good design is hard for a model, the company says, and editing one is harder, because the model has to work out what you want changed, how the design is structured, and what has to stay exactly where it is. Get that last part wrong and a simple request cascades into changes nobody asked for.

Brush handles it by treating design as a process rather than a set of isolated decisions. It can keep evaluating and improving its own output, which is how Canva describes moving a rough sketch toward something finished. Under the hood, the loop writes code, runs it against the design, looks at the page, and decides what to do next.

Everything it produces stays editable. Text, imagery, colours, layout, and the smaller details all remain available to the person who inherits the file.

What it was trained on, and what that implies

Canva started from an open-weights model it hasn't named, then post-trained it on actual design and editing work. The training signal combined design expertise with anonymised behaviour from users who opted in to improve Canva AI. Large-scale reinforcement learning let the model take on full assignments, learn from which outcomes worked, and adjust.

That training data is the part competitors can't buy. Vladimir Beskorovainyi, an enterprise AI architect who founded Besk Tech, told The New Stack that Canva isn't trying to out-build frontier labs. It picked a bounded problem, and bounded problems are where small specialised models have been beating large general ones for years. The moat sits in the artefact, he argued, because Canva can train on how real designs are actually constructed. A frontier lab can match the compute. Matching that training signal means owning the designs.

Rodion Sorokin, co-founder of the deep-tech venture studio The Capital Index, put the same idea more sharply: frontier models know the internet, Canva knows what hundreds of millions of people actually do with a slide deck.

Craft is the one with the leaderboard position

Brush has no public benchmark score. Craft does, and it's worth looking at because it sets the context Canva is operating in.

On DesignArena, a crowdsourced comparison platform, Craft sits fourth with an Elo of 1341 across 1,303 battles, a 66.7 percent win rate, and an average of 54.0 seconds per generation. The three models above it all come from OpenAI: GPT-Image-2.5 Sunburst at 1415, GPT-Image-2.5 Flare at 1387, and GPT Image 2 at 1362.

Canva's headline claim was that its research arm ranked as the world's second most advanced AI lab for image generation. That is a ranking of labs, not models, and the distinction matters. As a model, Craft is fourth, and its Elo rests on roughly 1,300 votes against Sunburst's 14,177. Expect the position to move.

The category breakdown is more useful than the overall rank. Canva says Craft ranks first of 87 models for People and Portraits and first of 85 for Landscapes and Nature. On Marketing Materials, the category a lot of Canva's customers actually work in, it ranks sixth of 85. On Cartoons it's 18th of 87.

Canva's own architecture explains some of this. Craft works in two stages: it builds a plan for the design first, settling element placement, relative scale, visual hierarchy and what text should appear, and only then renders, at up to 2K across more than 30 aspect ratios. Training came from Canva's template library, so posters, social posts, presentations and menus rather than photographs. That is why text stays legible and why a plan-first model does well when a title has to sit across the top of a nature scene.

The gaps Canva already named

Brush shipped without Brand Kit support, without the ability to reference attachments, and without better handling of image-heavy layouts. Canva says all three are in progress. For a marketing team, the first one is the blocker: until Brush understands brand colours, fonts and logos, every edit needs a manual check afterwards.

Neither model has its own price. Both draw on the AI allowance bundled into a Canva plan. Free accounts get limited access, Pro runs $144 a year for one person with ten times the AI of the free tier, Business runs $250 a year per person with twenty times.

Why the editing model is the one to watch

The pattern here isn't unique to Canva. Every platform that owns a workflow is now shipping a model that operates inside that workflow rather than competing at the frontier. Canva owns the design file, so it can train on edits. Adobe owns the timeline, so it can train on clips. The model matters less than the file format it can reach into.

There's a cost argument underneath the quality argument. A design agent that regenerates a whole slide burns tokens on the parts nobody complained about. An agent that patches one property spends a fraction of that, and the 46 percent figure is really a statement about how much of a typical edit was wasted work. If that holds outside a launch post, the economics shift for anyone running hundreds of edits a month rather than hundreds of generations.

For anyone building a content pipeline, the practical takeaway is smaller and more concrete. Generation has become cheap and interchangeable. Editing is where the hours go, and it's where a model that writes code and checks its own work has an advantage over one that redraws the whole frame.

Brush is a first version. Canva says so directly. But the direction is the one that pays.

Related articles