Canva Brush: Canva's Design-Editing VLM You Can Only Use Inside Canva

Canva Brush is the rare model launch where the headline number is a bill rather than a score. Canva unveiled it on October 29, 2026 — a vision-language model built by Canva Research to edit designs rather than generate them, live that day for Canva Pro subscribers inside Canva AI. The figures Canva leads with are 69% faster editing and 46% lower cost per edit task, both measured against Canva's own previous design agent. Quality, the company says, matched or exceeded that previous setup on most tasks. There is no leaderboard entry to check either claim against, because the board Canva's first in-house model landed on — Canva Craft, fourth on DesignArena's image-generation rankings one week earlier — ranks image generation, and Brush does not generate images. It rewrites the ones you already have.

That makes this a different kind of release from the one Canva shipped seven days ago, and worth reading differently. Craft was a showpiece: a model trained from scratch, pointed at a public benchmark, ranked against GPT-Image-2.5 Sunburst and GPT Image 2. Brush is an operating expense. It exists because Canva's AI bill got large enough to force a public admission — the company cut its full-year revenue growth guidance from about 30% to about 20% in August, with chief executive Melanie Perkins telling shareholders the company had been "relying too heavily on frontier models." Brush is the second step in that correction, and the interesting question is not whether it beats anything on DesignArena. It is what a 46% cut buys, and who gets to use it.

What Brush actually does

The mechanics are more specific than the press coverage suggests, and they are the reason this is not simply "an AI that edits slides." Brush reads a design's layout, layers and visual style, then acts on it through an agentic loop: it writes and runs code that adjusts design elements, inspects what actually happened on the canvas, and uses that observation to decide the next operation. Canva's own description is that it evaluates and improves its work rather than treating design as a series of isolated decisions.

Everything it produces stays editable. Text, imagery, colours and layout remain ordinary Canva elements you can drag, resize and retype afterwards. That is the structural difference from an image model: the output is a set of operations on Canva's design document, not a flattened picture, which is also why the model needs a canvas it can inspect rather than a prompt it can answer in one shot.

The tasks Canva names are the ones that eat time in real workflows rather than the ones that demo well:

• Restyle a design — change the colour scheme, the mood, the weather in a scene, without rebuilding the layout underneath.

• Sweep a multi-page document — rename every section header in a deck from one instruction instead of paging through.

• Reformat for a channel — turn a presentation slide into an Instagram graphic, keeping hierarchy and typography intact.

Training is described in the same register as Craft's but with a different reward. Brush started from "a leading open-weights model" — Canva does not name it, and no coverage we found corroborates a guess — then was post-trained on real design and editing tasks and put through large-scale reinforcement learning. The reward function scores visual quality, adherence to design principles and task completion, and it is calibrated by Canva's own professional designers. The training mix includes anonymised data from users who opted in to improve Canva AI.

The 69% and the 46%, read carefully

Both numbers come from Canva's internal testing, not from an external evaluator, and both share one undefined term: "Canva's previous design agent." That baseline is the whole comparison, and Canva does not say publicly what powered it. If it was a frontier model rented per task, a 46% cost reduction is a modest first bite out of a bill the company says it has already cut by nearly 90% since April. If the previous agent was already partly in-house, 46% is a large marginal gain on an already-cheap base. The launch post does not distinguish between those worlds, and neither figure can be reconciled with the "30 times cheaper than frontier alternatives" claim Canva made about its in-house image model in August, because the two are measured against different things.

What is verifiable is the surrounding programme. Perkins told shareholders in August that Canva had cut the cost of serving a single AI task by nearly 90% since April, and that its in-house style-transfer model runs 23 times cheaper than frontier alternatives, its image model 30 times cheaper and its video model 17 times cheaper. She also said Canva now routes work between its own models and outside labs depending on the task, and reuses its library of more than 165 million templates and assets "wherever possible before generating new content." Brush's 46% is the same argument applied to editing, at a company that says its users create more than a billion designs a month, where "a variation of even a few cents makes an enormous difference."

The speed claim has the same shape and the same caveats. "Up to 69%" is a ceiling across general design editing tasks, not a median, and Canva publishes no sample size, no task mix and no hardware. A vendor ceiling on an internal harness is a directional statement: the model is meaningfully faster than what it replaced, by some amount, on the tasks Canva chose to measure. It is not a number you can put in a capacity plan.

There is also no price to check. Brush is not sold by the token, by the image or by the call. It is a feature of a Canva Pro subscription, and Canva has not published a credit cost per edit — which means the 46% saving lands on Canva's gross margin, not on a line item a customer can see. If you are trying to decide whether design editing is worth paying for, the only price signal available is the Pro subscription itself.

The number you cannot check at all

Canva's quality claim — matched or exceeded the previous agent on most editing tasks, strongest on restyling and format adaptation — rests entirely on an evaluation Canva designed, scored against a definition of "good" set by Canva's designers. That is a legitimate way to build a product and a weak way to convince an outsider. There is no public sample size, no published rubric, no third-party run, and no arena for design editing of the kind DesignArena runs for image generation. Canva's own description of the reward system is the most detail anyone has: visual quality, design principles, task completion, hundreds of thousands of simulated interactions.

The contrast with Craft is instructive. Canva's Craft post, and our own write-up of it, recorded a debut at fourth place on DesignArena's image leaderboard on October 22, behind three OpenAI models. That board ranks image generation from a text prompt, and read on October 30 it still carried Canva Craft in fourth place — a reminder that arena ratings move with votes and are a snapshot, not a fixed fact. What matters here is the absence: there is no row for Brush, and no equivalent board anywhere else. Brush edits existing designs rather than generating new images, so the arena has nothing to score it on, and Canva's quality claim for Brush has no counterpart to the crowdsourced, adversarial signal Craft already has.

This is not a reason to dismiss the model. It is a reason to be precise about what has been demonstrated. Craft has an external, adversarial, crowdsourced signal — which is why its rank can be quoted and argued about. Brush has a product claim: fewer seconds, smaller bill, no visible drop in quality. Two of those three are measurable by Canva and one is a matter of taste, and Canva has published evidence for none of them in a form anyone else can rerun.

Who can use it, and on what terms

Availability is narrow and specific. Brush powers "supported design edits" in Canva AI for people on Canva Pro plans, with wider availability described only as "coming soon." To make sure it is handling the edit, you open a design in the editor or link one in Canva AI and select "Brush" from the model dropdown. Canva's launch post lists no timeline beyond that, and it lists three things still missing from the first version: Brand Kit support, the ability to reference attachments, and better handling of image-heavy layouts.

What the post does not list at all is a developer surface. Canva Craft arrived with a route into Canva's public API — preview-gated, as we found on launch day, but documented and callable. Brush's announcement describes an in-product feature and nothing else; no endpoint, model identifier or price appears in the launch materials or in the developer documentation we checked. What Canva does document is a Design Editing API, which lets apps read and edit the "ingredients" of a design — pages and elements — from inside the editor, still flagged by Canva as under active development. That is an Apps SDK surface for building your own app inside Canva's canvas, not an endpoint that calls Brush. So the practical answer for anyone outside Canva is that Brush is not a model you can call. It is a capability inside a subscription, and the subscription is the interface.

That is a deliberate trade, and Canva is fairly open about it. The company's stated direction is "bringing together the best of leading partner models with our own specialised models built for the creative tasks at the heart of Canva." In-house models exist to protect margin on the tasks Canva can serve itself, while partner models keep handling everything else. Brush is not a replacement for the frontier models Canva still rents. It is a substitution on one high-volume task, and Canva is gearing up to unveil "the next generation of Canva AI" with the savings folded in.

If you are building this yourself

The useful lesson in Brush is not the model. It is the loop around it. A design-editing agent needs four things: a document model it can read and write at element level, a way to execute operations and see the result, a reward signal that encodes what "good" means for your domain, and a vision-language model strong enough to follow the instruction without touching what it was told to leave alone. Canva has all four in-house, and three of them predate Brush. The base model is the commodity piece — Canva says so itself by starting from open weights and spending its effort on post-training and reward design.

That is the part a smaller team can actually copy. An open vision-language model plus your own document schema plus an execution-and-inspection loop gets you a prototype long before you have Canva's design taste or Canva's data. The model side is where a router helps, and it is worth being exact about what we do here: OrcaRouter does not serve Canva Brush — it is a proprietary Canva feature with no public endpoint, and we do not host it or imply otherwise. What we offer is the other half of a build like this, one API for 200+ models at provider list price with 0% markup, so the vision models you benchmark against each other arrive on the same key, and a vendor price cut reaches you the same day it lands. Automatic failover is what lets you put an unproven model behind a real workload without betting the product on it, and the routing DSL lets you compose several models into one call when you want a cheap model to draft and an expensive one to verify the result.

For a team choosing a design-editing stack in late 2026, the honest comparison is not Brush versus some open model. It is "pay Canva per seat and accept its taste" versus "own the loop and rent the model." Brush tells you which half of that problem Canva decided was worth owning, and it is not the weights.

What would change the picture

Four things are worth watching, and none of them is a benchmark score. First, the rollout: Brush is Pro-only today, and Canva has said nothing about when it reaches free accounts or which plan tier pays for it. Second, Brand Kit support, which is the difference between a model that restyles one design and one that applies a brand system across a hundred — the capability that would make an agency workflow actually change. Third, an API: if Canva exposes Brush the way it exposed Craft, the model becomes evaluable by outsiders for the first time, and the 69% and 46% get something to be checked against. Fourth, the editing arena that does not exist yet: if DesignArena or anyone else starts ranking design-editing agents the way it ranks image models, Brush would get the external signal Craft already has.

Until one of those lands, the accurate summary is this. Canva Brush is a real model doing a genuinely harder job than image generation — reading structure, preserving what should not change, and leaving the output editable — and it is the second proof point in Canva's campaign to stop renting the most expensive part of its AI stack. The efficiency claims are plausible, internally measured and unverifiable from outside. The quality claim is Canva's designers grading Canva's model. And the model itself, for all that it is the most interesting thing Canva Research has shipped this month, is not something you can call, host, benchmark or buy separately. It ships inside Canva, and it is priced that way.