Canva Craft Review: #4 on DesignArena, Text in 55 Languages, and What It Actually Costs

On October 22, 2026, Canva shipped its first image generation model built entirely in-house. Craft, trained from scratch by Canva Research, went live for every Canva user globally in Dream Lab and through the public API the same day. The headline numbers are strong — a #4 debut overall on DesignArena, #1 for people and portraits — but Canva published no price, no technical paper, and the rankings are its own reporting of a crowdsourced leaderboard. Here is what is verified, and what is marketing.

Quick answer

• What it is: Canva's first in-house image generation model, trained from scratch by Canva Research and launched October 22, 2026. Training emphasizes design over photographs — Canva says the data includes millions of human-made templates from its library.
• The rank: #4 overall debut on DesignArena per Canva; #1 for People & Portraits (of 87 models) and Landscapes & Nature (of 85 models), #2 for Abstract & Patterns, #4 for Products. Company-reported crowdsourced rankings, not independent benchmarks.
• How to get it: all users in Dream Lab now; developers pin craft_v1_2k in the public Connect API. Support inside ChatGPT and Gemini via Canva's MCP is planned, not live.
• The pricing catch: no published per-image price, and Canva hasn't said which AI-allowance tier Craft draws from. Craft-fast and Craft-ultrafast variants promise "a fraction of the cost of leading frontier models" — no dates or numbers.

Why "trained on templates" is the whole bet

Nearly every image model you have used was trained primarily on photographs — excellent for a portrait, a poor way to learn what designers get paid for: where the headline sits relative to the logo, how a price fits inside a menu column, why one layout reads as premium and another as clutter. Canva's core claim is that Craft's training data includes millions of human-made templates from its own library: posters, social posts, presentations, infographics, menus, and typography-heavy layouts.

Unlike flat images, templates carry structure — layer boundaries, alignment, spacing, typographic hierarchy. A model that trains on them learns composition as a first-class skill rather than hoping a single diffusion pass gets it right. Whether this actually produces better design work than the photo-trained frontier models is the question independent testing must settle — and as of launch, there are no published third-party tests, only Canva's own rankings. The business logic is equally clear: Dream Lab previously ran on Phoenix, the model from Leonardo.ai, the start-up Canva acquired in 2024. Owning the model means Canva stops paying other companies' API bills and controls quality and speed itself.

The two-stage architecture, in plain terms

Most image generators turn a prompt straight into pixels in one pass. Craft splits the job: a planning model interprets the prompt and writes a composition plan — elements, spatial relationships, relative scale, visual hierarchy, text placement — then a rendering model turns that plan into an image at up to 2K resolution across 30+ aspect ratios.

That separation powers Canva's two headline features. First, legible text: Craft renders text into images in 55 languages, per Canva, because text is a layout object with a position and size in the plan, not texture painted in the render pass. Second, targeted edits: an explicit composition means you can change one element without unnecessarily altering everything else — and Canva's Magic Layers can convert outputs into editable design elements via JSON boundary boxing.

After pre-training, Canva optimized Craft with supervised fine-tuning and reinforcement learning from its professional designers — over 25,000 image pairs assessed head-to-head against leading models, with the designers' judgments winning when they disagreed with automated scores. Company-run testing, not independent review, but an interesting stated philosophy.

Scale also shaped the engineering. With more than a quarter of a billion people using Canva — the company says users create over 400 designs every second — Canva used advanced distillation to reduce Craft's inference steps "to a fraction" of the original model, and built a custom inference engine for the planning stage it says is an order of magnitude faster than its first internal version.

What the DesignArena ranking actually proves

Canva's loudest stat is Craft's #4 debut on DesignArena, which makes Canva Research, in its words, "the world's second most advanced AI lab for image generation" behind OpenAI. The subcategories are strong: #1 for People & Portraits (of 87) and Landscapes & Nature (of 85), #2 for Abstract & Patterns, #4 for Products.

Keep three things in mind. First, DesignArena is a crowdsourced arena — votes from real users, not a controlled benchmark — and Canva cites its own ranking, which no independent outlet has separately confirmed. Second, a leaderboard debut measures first impressions, not sustained production quality. Third, Craft arrives into a crowded field: Canva names Google, OpenAI, ByteDance and Meta as the competitors it built Craft to face. Debuting #4 in that company is a real signal — but about a launch-week leaderboard, not a verdict on your next campaign.

If you evaluate image models for a living, our guide to free AI image generators covers what the rival tools' free tiers actually allow, and our FLUX 3 Max review shows how one recent launch-week pricing deal was structured.

Craft vs. the competition

Craft's genuine differentiator in this table is the training corpus: the only model here built on millions of human-designed templates rather than photographs. Its weakest cell is the cost model — every other model publishes a way to pay; Craft publishes an allowance. For the bigger pricing picture across model categories, see our Gemini 4 Argon pricing breakdown.

How to access Craft (and how the API works)

For users, Craft is live in Dream Lab — open it from the bottom-left panel on the Canva homepage. Canva's Dream Lab help page says some features use your AI allowance and that you can generate up to four images per prompt. For developers, Craft is in the Connect API image-generation endpoint per Canva's developer community: pin the model with craft_v1_2k on POST /v1/image-generations, then poll GET /v1/image-generations/{jobId} until the job succeeds. The download URL is valid for 24 hours, or set asset_upload.type to upload to save the image to the user's uploads. The call needs a user access token with the asset:write scope, and prompts can be written in any language.

One nuance: Canva's newsroom post says Craft is available via the public API "and the Canva MCP Server," then immediately labels the MCP section "Coming Soon." Third-party chatbot support for ChatGPT and Gemini is a roadmap item, not a live feature — do not buy or build around it until it ships.

The pricing fine print, honestly stated

The allowance system, from Canva's pages and launch-day reporting: Free gets a small monthly allowance (reported at up to 20 Standard or Premium uses), Pro a larger one (reported at up to 200 Premium uses plus a smaller Ultra allowance), Business more again. Plan prices vary by region — check your local page rather than quoting one market's numbers as global.

The reporter math: if Craft draws from the Premium tier, Pro works out to roughly six cents per image and Business to roughly five cents — reporter arithmetic, not Canva's, and if it draws from the Standard tier, paid plans may pay effectively nothing within fair use. Allowances reset monthly (on your billing date for paid plans). For API builders, jobs consume the end user's credits — your cost is your customer's plan, not a line item you can forecast. As for the promised Craft-fast and Craft-ultrafast at "a fraction of the cost of leading frontier models": no dates, no numbers.

Pros and cons

What Craft does well

• Design-first training data. Millions of human-made templates is a genuinely different starting point from the photo-trained field — the right bet for posters, menus, and social graphics where layout carries meaning.
• Two-stage planning. Plan-then-render directly attacks the failures that matter in design: illegible text, misplaced objects, edits that break the rest of the image.
• Text in 55 languages. If independent testing backs this up, multilingual typography is a real differentiator for global marketing teams. (Company claim — verification pending.)
• Editable outputs. JSON boundary boxing feeding Magic Layers turns generated images into editable design elements instead of flat pictures.
• Complete day-one rollout. All users globally in Dream Lab plus a public API at launch is unusually complete.

What to watch

• No published price. No per-image rate and no confirmed allowance tier — commercial planning runs on reporter arithmetic, not vendor data.
• Rankings are company-reported. #4 overall, #1 in portraits and landscapes — Canva's reading of a crowdsourced leaderboard, with no independent confirmation yet.
• No technical paper. The architecture is described in outline, not in publishable detail.
• MCP and chatbot support are roadmap, not reality. Do not buy or build around ChatGPT/Gemini integration until it ships.
• Vague training provenance. "Millions of human-made templates" tells you the format, not the licensing or artist-compensation details designers will ask about.

FAQ

Is Canva Craft free?

There is no dedicated price for Craft. It is available to all Canva users in Dream Lab and draws from Canva's AI allowance tiers — Canva hasn't said which tier it counts against. Free plans get a small monthly allowance; Pro and Business get larger ones.

Can developers use Craft through an API?

Yes. Canva's developer community confirms Craft is live in the Connect API image-generation endpoint: send POST /v1/image-generations with model: craft_v1_2k, poll the job until it succeeds, and collect the download URL. You need a user access token with the asset:write scope. Third-party chatbot integration via Canva's MCP server is planned but not yet live.

How good is Craft at text in images?

Canva says Craft can render text in 55 languages, and its two-stage design specifically targets legibility — text is planned as a layout object before rendering. That is the company's claim; independent tests have not yet been published.

Can I use Craft inside ChatGPT or Gemini?

Not yet. Canva plans Craft support inside third-party chatbots through its MCP server, but the newsroom post labels the MCP as "Coming Soon." Until then, access is Dream Lab for users and the public API for developers.

What did Craft replace in Dream Lab?

Dream Lab launched in October 2024 running on Phoenix, the model from Leonardo.ai — the Sydney start-up Canva acquired in 2024, whose researchers also contributed to Craft. Craft is Canva's first model built and trained entirely by its own research lab.

Key takeaways

• Canva Craft, launched October 22, 2026, is Canva Research's first image generation model trained from scratch — on millions of human-made design templates rather than photographs.
• Its two-stage architecture plans the composition first, then renders — a direct attack on the illegible-text and broken-layout failures that plague design work.
• Canva reports a #4 overall debut on DesignArena with #1 finishes for People & Portraits and Landscapes & Nature — crowdsourced, company-reported, and not yet independently confirmed.
• It is live for all users in Dream Lab and through the public API (model craft_v1_2k); ChatGPT/Gemini support via MCP is roadmap, not reality.
• There is no published per-image price and no confirmed allowance tier — plan around a subscription allowance until Canva publishes a rate card.

Published October 23, 2026. The pricing situation and the MCP rollout are the two things to re-check before building anything commercial on Craft — we will update this analysis when Canva publishes numbers.