Canva shipped two models in eight days.
Canva Craft on 22 October, which generates images. Canva Brush on 29 October, which edits designs.
The split is deliberate and Canva says so plainly: "Craft generates the image, and Brush refines the design around it."
That sentence is the whole thesis. Generation was never the hard part. The hard part is what happens after, when someone looks at the thing and says "make the headline smaller and fix the spacing."
That is the job almost nobody wants. And it is the job Canva just built a model for.
What actually shipped
Brush is a vision-language model for design, built by Canva Research, their 160-person AI lab.
It is live now for Canva Pro subscribers, inside Canva AI. You open a design, pick Brush from a dropdown, and describe the change.
| Claim | Detail |
|---|---|
| Speed | 69% faster on general editing |
| Cost | 46% lower per edit task |
| Basis | Canva's previous design agent |
| Method | Not published |
| Access | Canva Pro, wider rollout later |
Both numbers are Canva's own, against their own previous agent, with no test method published and no outside benchmark.
i want to be fair about that. Comparing yourself to your own last version is the only honest comparison at launch. But it means you cannot tell whether Brush is good in absolute terms or just better than something that was already there.
The interesting mechanism
here's a question people always ask about AI design tools: how does it know what to change?
Brush does not treat a design as a pile of separate decisions. It runs an agentic loop. It writes code, runs that code to adjust design elements, then inspects what is actually on the page and decides its next step.
That is a meaningful difference from a single-shot generation. It can check its own work mid-task.
And the output is a normal Canva design. Every text box, image, colour and position stays editable. You can take over by hand at any point.
That last part matters more than the speed number. A design you cannot edit is a picture. A design you can edit is a file you can keep working in.
The part that is actually hard
Most tutorials tell you the model's job is to make the change you asked for.
It is not. The job is to make the change and leave everything else alone.
Canva says this directly in their post. An editing model has to understand three things: what you want changed, how the design is built, and what must stay exactly the same.
Get the third one wrong and a small request becomes a pile of changes nobody asked for.
i have done this to myself with design tools for years. You ask for one font change and come back to a document where the margins moved, two images resized, and a heading is now a different weight.
Nothing is broken. Everything is slightly wrong.
That is the failure mode that kills these tools in real use. Not bad taste. Over-editing.
Canva's sample prompt shows how much hides inside one request:
"make me a 3-page pitch deck for my bakery with art nouveau visuals and backgrounds"
To do that well the model has to plan a story across three pages, write and place text, choose images, and hold fonts, colour and spacing consistent from page one to page three.
Three pages is where it gets hard. One page can be pretty by accident. Three consistent pages is a system.
Why Canva built its own model
what actually happens is that the quality argument is a decoy and the cost argument is the real one.
Canva's post says general-purpose models can reason through design edits fine. They just cost more to run.
Canva has hundreds of millions of users making billions of designs. A few extra cents per edit task is not a rounding error at that scale. It is the difference between offering AI editing to everyone and offering it to a small slice.
So they trained a specialised model that is cheaper per task.
46% lower cost per edit is not a feature. It is the business model.
That is the most underreported line in the launch. Every article covered the 69%. Almost none led with the 46%.
i think that is backwards. The speed number is for the demo. The cost number is why the product exists.
What it is built on
Brush started from "a leading open-weights model." Canva does not say which one.
They post-trained it on real design and editing tasks, using in-house design expertise plus anonymised data from people who opted in.
Then large-scale reinforcement learning. The model worked through complete design tasks and learned from what scored well, using a reward system that grades visual quality, how well the layout follows design principles, and whether the task was actually completed.
Canva notes that the definition of good design was set by professional designers on their team.
That is the right way to do it and also the fragile part. The model inherits whatever those designers consider good. If your taste differs, you are arguing with a rubric you cannot see.
A detour about the word brush
A brush paints.
Brush is not a painting tool. It does not put pigment anywhere. It edits layouts, restyles images, and reflows text across pages.
The name describes the old metaphor, not the new product. A brush makes a mark. This thing removes marks, moves them, and decides which ones should stay.
i keep thinking about how many design tools are named after physical gestures. Brush, pen, lasso, wand, stamp. All of them borrowed from the era when your hand was the input.
An agentic loop has no gesture. It has a plan and a check.
Canva picked a name that makes the product feel familiar. It also makes it harder to explain, because the thing it names is the one job the model does not do.
Who should not use this yet
Brand Kit support is missing.
Canva says it is coming, along with attachment references and better handling of image-heavy layouts. Until it lands, if your work depends on exact brand colours and licensed fonts, Brush is a risk. It will restyle things and you will spend the time fixing them back.
If you are on Free, you cannot use it. Pro only, for now.
If your design work is genuinely custom, skip it. Illustration, packaging, anything where the layout is the idea rather than the container.
And if you are a designer who bills for production edits, this is the part of the job that just got cheaper for everyone. That is worth thinking about honestly rather than optimistically.
The people who should try it are the ones who make the same edit forty times. Resizing a deck for four channels. Updating a date across twelve slides.
Restyling a template for a new season.
That is the work. It is not creative and it is not going away.
The bit that stuck
Canva's post has a section titled "Great design happens in the edit."
i read that twice. It is a strange thing for a company whose product is templates to say out loud.
Templates were always the thing you started from. The edit was the friction between the template and your actual content.
Brush is Canva admitting that the friction was the product all along.
The template got everyone in. The edit is what keeps them.
If the 46% cost claim holds at their scale, and the over-editing problem is genuinely solved rather than demoed, then the boring half of design work just got automated by the company that already owned the boring half.
That is a much bigger deal than 69% faster. And almost nobody wrote about it.
Top comments (0)