Canva shipped its own image model on October 22. It is called Craft, it sits in Dream Lab for every Canva user, and it is the first model Canva has built and trained from scratch rather than rented from a partner. If your brand or social team lives in Canva, it is already in the dropdown they open every morning.
This is a docs-based review. I have not run a campaign through Craft. Everything below comes from Canva’s own launch post, developer documentation, help center and AI terms, checked against independent coverage, and I will tell you where those sources stop.
The short version: Craft is a genuinely useful model with the right architecture for marketing work. It was also trained on the same template library your competitors pull from, Canva’s own terms say your output may not be unique, and the meter that pays for it is a personal allowance that nobody on your team can see in aggregate. Those three facts decide how you should use it, more than any leaderboard does.
What Craft is
Most image models go from prompt straight to pixels. Craft splits the job. A planning model reads the prompt and writes out a composition: which elements appear, where they sit, how big each one is, the visual hierarchy, and any text that has to show up. A second model renders that plan at up to 2K resolution in more than 30 aspect ratios. Canva says this is why Craft can put legible text where you asked for it and change one element without regenerating everything else. Canva’s announcement spells out the architecture; The Verge and Betanews report the same two stages.
The training data is the part marketers should read twice. Canva says Craft’s training “includes millions of human-made templates from Canva’s content library, spanning professionally designed posters, social posts, presentations, infographics, menus and typography-rich layouts.” Designers then judged more than 25,000 head-to-head comparisons, and “when automated evaluations and human judgments differed, we sided with the designer.”
You reach it three ways. In the editor through Dream Lab. Through the Connect API by pinning the model value craft_v1_2k. And, later, through the Canva MCP server, which Canva lists as coming soon, so an AI assistant can make a launch graphic as one step in a bigger task.
What changed
Until now, Canva’s image generation leaned largely on partner models. Stef Corazza, who runs Canva Research, explained the reason for building in-house in plain money terms: across more than a billion designs a month, “even a variation of a few cents per generation makes an enormous difference.” Craft is a quality play and a margin play at the same time. Two cheaper variants, Craft-fast and Craft-ultrafast, are announced with no prices and no dates.
On quality, Canva says Craft debuted fourth on DesignArena, a crowdsourced leaderboard where people vote between anonymous outputs, behind three OpenAI models. Count labs instead of models and you get Canva’s “second most advanced AI lab” line. Canva says Craft ranks first of 87 models for People and Portraits and first of 85 for Landscapes and Nature. Betanews noted the score had already slipped from 1340 to 1330 a day after launch, which is what crowdsourced leaderboards do.
One detail from the category chart is worth more than the headline rank. An independent review that read Canva’s own chart puts Craft at sixth of 85 for Marketing Materials. The model trained hardest on marketing templates wins at portraits and landscapes, and sits mid-pack on the exact category your team will use it for. Do not overread one leaderboard. Do stop treating “built for design” as a promise that the poster comes out right.
What works
The plan-first architecture is the right idea for marketing assets. The costliest AI poster failure is not ugliness, it is a misspelled offer or a drifted date. A model that decides the text before it draws the letters attacks that failure directly, and Canva’s text support covers 55 languages, which matters for field and regional teams who currently send everything back to a central designer.
The image lands where the work happens. For a Canva-native team, the real saving is not the generation, it is skipping the export, upload, resize and reposition loop. Craft outputs are built to be split by Magic Layers into editable elements, so a generated headline can become an editable element a human corrects instead of a baked-in string someone has to paint over.
The aspect ratio range is practical in the editor: the 4:5 feed version, the 9:16 story and the 16:9 banner can start from the same prompt instead of a crop. The API is narrower and exposes six ratios.
And the developer surface is honest about itself. The API reference documents a 20 requests per minute limit per user, a 429 with credit_quota_exceeded when the user runs dry, a separate cooldown error after heavy use, and a model_not_available failure when Craft is pinned and busy. Canva also states that when you pin a model, it “never uses a different model to serve the request.” That is exactly the behavior you want if a brand team has approved one model’s output style.
What breaks
Sameness is the first problem, and it is structural. Craft learned layout, hierarchy and type from Canva’s template library. That library is also what millions of other brands, including your direct competitors, start from every day. Canva’s AI Product Terms say it out loud: “outputs may not be unique, and other users may receive similar Outputs from our AI Products. Your ownership above does not extend to other users’ Outputs.” You own your output. You do not own the look. A promo graphic generated from a short prompt is likely to land near the center of what Canva templates already look like, which is the opposite of what a brand is paid to do.
Text is better, not finished. Canva’s own launch post says the team will keep improving “things like fine print and technical drawings.” Fine print is where marketing lives: prices, dates, terms, disclaimers, URLs. The terms add that Canva “has not verified the accuracy of the Output” and recommends independent advice before you use any output in a commercial context. Read that as: a human proofs every number before it ships, with no exceptions for the social team on a Friday.
Indemnity is narrow and its scope is unclear. Canva Shield says eligible Enterprise customers are indemnified for AI output at no extra cost when using “Magic Studio products.” Pro, Teams and Business seats get no such cover. And none of the launch material I read says in so many words whether Craft in Dream Lab counts as a Magic Studio product. If you are on Enterprise, get that confirmed in writing before Craft images go into paid media.
The API puts the bill on the person, not the product. Every API job needs a user access token with the asset:write scope and “consumes the user’s AI credit allowance.” The download link expires after 24 hours. The endpoint is a preview, so breaking changes can arrive without a new API version, and public apps built on it cannot pass Canva’s app review. Canva’s developer forum post says Craft jobs take roughly 50 to 60 seconds end to end. That rules out anything a customer waits on, and it means a vendor tool or internal script that “uses Craft” is quietly spending one named employee’s monthly allowance.
There are compliance strings attached too. The same terms prohibit using AI output to mislead anyone that it was made by a human, and prohibit removing provenance metadata such as C2PA tags. If your export or ad ops process strips metadata by default, check it before Craft images flow through it.
Who it is for
Craft is for Canva-native teams who make a lot of supporting assets: social posts, event graphics, regional variants, sales one-pagers. For them, a decent image inside the editor that splits into layers saves real hours.
It is a weaker fit for the hero work that carries a brand: the campaign key visual, the product shot that runs in paid, anything where distinctiveness is the point. Not because the pixels are bad, but because a model trained on the shared library pulls toward the shared look, and the terms tell you other users may get something similar.
It is not yet a fit for product teams building their own generation feature. User-paid credits, a preview API and a minute per image make it an account integration, not an engine.
Pricing and limits
There is no separate Craft price. Generation draws on the shared monthly AI allowance that comes with each Canva plan. Canva’s help center article Understanding your AI usage lists “Create an image with Canva AI” and Magic Layers as Premium AI tools. Canva has not said which tier Craft itself bills at, so treat the Premium column as the closest guide, not a promise.
The allowances, per the help center: Canva Free and Pro Lite get up to 20 uses of Standard or Premium AI tools. Pro, Teams and Nonprofits get up to 200 Premium uses or 20 Ultra uses. Business and Enterprise get up to 400 Premium or 40 Ultra. Pro and Business trials get 20, the same as Free. These are ceilings that assume simple prompts and one tool type; Canva says complex prompts and mixed tools land at the lower end.
Three details matter. First, the allowance is per member: “Allowances aren’t shared or pooled across the team or organization.” Your designer who generates all day will hit the wall while the rest of the team’s allowance sits unused, and you cannot move it across. Second, splitting a Craft image into layers is also a Premium use, so the generate and edit loop costs twice. Third, the terms call usage limits “operational controls and not fixed entitlements” that Canva may change. AI Pass, the add-on for heavier users, is sold as up to 40 times Pro’s allowance or 20 times Business’s, without a public dollar figure on the page.
Usage is visible per person under Settings, Billing. I could not find an admin view that rolls it up by team.
What to do this week
Run a sameness test before anyone falls in love. Have two people on the team write the same plain prompt for your most common asset type, say a sale announcement for Instagram, and generate in separate accounts. Put the results next to the top three Canva templates for that use. If you cannot tell your brand apart from the template, your prompts need your brand in them: colors, type rules, composition rules, the things a competitor will not type.
Write a one-paragraph Craft rule into your brand guidelines. Supporting assets yes. Campaign key visuals and paid hero images only after a designer review. Every price, date, offer and legal line proofread by a person, and set as an editable text layer rather than left in the generated image where possible.
Map who generates the most. Ask the two or three heaviest Canva users to check their AI usage under Billing this week. If one person is already near the ceiling mid-cycle, that is your AI Pass conversation, priced against the hours saved, not against the feature list.
Ask your account rep two questions in writing if you are on Enterprise. Does Canva Shield indemnity cover Craft output generated in Dream Lab and through the API? And which AI tier does a Craft generation bill at? On Pro or Business there is no indemnity to ask about, which is its own answer for paid media.
Check your metadata path. Export one Craft image through your normal ad or social workflow and see whether provenance data survives. If your tools strip it, fix that before volume goes up.
Hold off on API builds. If someone wants to wire Craft into a content pipeline, wait until the API leaves preview, and make sure no script is running on a departed employee’s token.
Craft is the best reason yet to keep simple image work inside Canva. Just remember what it learned from, and who else learned from the same library.
