Canva launches Brush, an AI model that edits designs by writing code

Brush arrived a week after Canva Craft, the Australian design software company's first in-house image-generation model, which launched on October 22. Both were developed by Canva Research, the company's 160-person AI lab. According to Canva, Craft entered crowdsourced benchmarking platform Design Arena as its fourth-ranked model, behind three OpenAI models: GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and GPT Image 2.

In its launch blog, Canva said generating high-quality designs is one of the biggest challenges for AI models, but that "editing them is even more complex", because a model has to understand what a designer wants changed, how the design is structured and what needs to stay exactly as it is. Otherwise, a simple edit request can produce a "whole bunch of changes" nobody asked for.

Canva said it built Brush around that problem. "Rather than treating design as a series of isolated decisions, Brush can continually evaluate and improve its work," the company said, adding that the model "uses an agentic loop that writes and runs code to adjust design elements, and can then inspect what's happening on the page to guide its next step".

Brush began from an unspecified open-weights model that Canva then post-trained on real design and editing tasks, drawing on design expertise and anonymized insights from people who opted in to help improve Canva AI. "We combined post-training with large-scale reinforcement learning, giving the model opportunities to work through complete design and editing tasks, learn from what worked, and continually shape its future behavior," the company said.

Canva says Brush is up to 69% faster than its previous design agent at 46% lower cost per task. Those figures are self-reported and compare Brush with the company's earlier agent rather than with frontier models. Canva says everything Brush creates is fully editable, and that this is the first version, with work under way on reference attachments, better handling of image-heavy layouts and Brand Kit support.

Vladimir Beskorovainyi, an enterprise AI architect and founder of Besk Tech, told The New Stack that the question of whether Canva can challenge the frontier labs is "posed in the wrong frame". He said Canva "is not challenging frontier models; it is in fact declining to compete with them", because reading how a design is assembled and editing it is a bounded problem, and bounded problems are where small specialized models have quietly beaten large general ones for years. He argued that "the moat here is the artifact, not the weights", since Canva can train on how real designs are actually put together, a signal he said is harder for a frontier lab to match.

Dr. Morissa Schwartz, founder of Dr. Rissy's Writing & Marketing, told The New Stack she has not yet used Brush but knows how she would test it: whether a headline change leaves the price, disclaimer and layout intact. "Canva's opportunity is to make those small, frequent revisions dependable enough that a marketing team spends less time checking what else moved," she said. "A beautiful edit can still be an expensive proofreading assignment," she added.

Rodion Sorokin, co-founder and CTO of deep-tech venture studio The Capital Index, said "Canva isn't trying to out-muscle the frontier labs; it's out-niching them". He said Brush was trained with reinforcement learning against what Canva's own designers call good design, "a playbook any well-funded vertical player can now copy". "The clever bit is that Brush doesn't paint pixels: it reads the layers, writes small pieces of code to change them and checks its own work, which is how a designer edits, not how an image generator guesses," he said. He added that if Canva's speed and cost figures hold up outside a launch blog, "specialized models will win on unit economics, not just quality: generation becomes the commodity, and editing becomes the business".