Canva Craft Debuts Fourth on Design Arena, but Canva's Own API Sends Text Elsewhere

Canva Craft scored 1,341 on Design Arena's image leaderboard the day after it launched — fourth place, 65 points behind GPT-Image-2.5 Sunburst, and just under GPT-Image-2.5 Flare and GPT Image 2. Canva shipped it on October 22, 2026, to every Canva account through Dream Lab and to developers through a public API, and the company's framing is that a single fourth-place model makes it the second-strongest image lab in the world. The more useful detail sits one page deeper, in the reference for the endpoint that serves the model, where Canva lists which of its three image models to pick for a given job. For images that contain text, it points somewhere other than Craft.

That is not a scandal, and it is not a reason to ignore the release. Canva Craft is the company's first image model trained from scratch, it uses a genuinely different architecture from the prompt-to-pixels models it is being ranked against, and it is the first time Canva's design templates have been turned into training data at this scale. But the launch is being sold on two claims — legible type and precise layout — and neither is what Canva's own developer documentation promises you. What follows is the version of the release you get when you read the leaderboard, the API reference, and the business context together instead of the press release alone.

What actually shipped on October 22

Three things shipped together: a model, a surface, and an API.

• The model — Canva Craft, internally CRAFT-1 2K. Canva Research built it from scratch with researchers from Leonardo.ai, the image-generation studio Canva acquired in 2024.

• The surface — Dream Lab, the generation tab inside Canva's editor, available to all account tiers including Free, subject to each plan's usage allowance.

• The API — the craft_v1_2k value on Canva's image-generation endpoint, covered in detail below.

The architecture is the part worth understanding, because it explains both the strengths and the odd corners. Craft runs in two stages. The first stage reads the prompt and writes a plan: which elements exist, where they sit relative to each other, what their relative scale should be, what the visual hierarchy is, and what text needs to appear and where. The second stage renders pixels against that plan. Most image models collapse both jobs into one pass and hope the composition emerges; Canva's bet is that deciding the layout first produces better placement, more legible lettering, and — the practical payoff — edits that can target one element without redrawing the rest of the frame.

The training data follows the same logic. Craft was trained heavily on millions of human-made Canva templates — posters, social graphics, slide decks, menus, infographics, type-heavy layouts — rather than on photographs. Canva says it followed that with supervised fine-tuning and reinforcement learning against feedback from its own professional designers, and that those designers judged more than 25,000 head-to-head image pairs during development. Those figures are Canva's, and the comparisons behind them have not been published.

The headline specs: output up to 2K, more than 30 aspect ratios inside Canva, prompts accepted in any language, in-image text rendered in 55 languages, and output that can be converted into a layered, editable design through Magic Layers. Two cheaper, faster variants — Craft-fast and Craft-ultrafast — are promised but undated. An MCP server that would put Craft inside ChatGPT and Gemini is also promised and also undated.

The ranking, minus the press release

The claim Canva is leading with is that Craft debuted fourth on Design Arena's image board, behind three OpenAI models, and that this makes Canva the world's number-two image lab behind OpenAI. Both halves are worth separating.

The fourth place is real and publicly visible. On the live image leaderboard, the top of the board reads GPT-Image-2.5 Sunburst at 1,406, GPT-Image-2.5 Flare at 1,382, GPT Image 2 at 1,362, then Canva Craft at 1,341. Next down are Reve 2.1 at 1,321, MAI-Image-2.6 at 1,317, and Nano Banana 2.1 at 1,291. So the model is not merely present — it is comfortably clear of Google's, Microsoft's and xAI's current image models, which is a stronger result than the fourth-place framing suggests.

The lab ranking is a different kind of claim. It is arithmetic performed on a single model's position: one fourth place becomes "second-best lab," a statement about an organisation rather than about the thing you can call. Every entry on the public board is a model, and a lab with one strong entry and no second has not been measured the way OpenAI has.

Two caveats apply to the numbers themselves. First, these boards move. Between two captures of the same page roughly twenty minutes apart, GPT Image 2 read 1,382 and then 1,362 — a 20-point swing in third place with no announcement in between. Second, the board does not display vote counts on the chart, so there is no way to tell from the page how settled a position is. Canva itself describes the placement as preliminary.

The subcategory claims are the ones to hold loosely. Canva's developer announcement says Craft ranks first in People & Portraits and in Landscapes & Nature. On the public board's subcategory tabs at the time of writing, Craft ranked second in Abstract, eight points behind Sunburst, and the tab set is named differently from Canva's phrasing — Landscape and Portrait rather than Landscapes & Nature and People & Portraits. The claim may well be true on the tabs Canva measured; it is not something a reader can confirm from the public page today.

What is missing from every version of the ranking story is a hands-on test. No outlet has published side-by-side outputs at the time of writing, and Canva has not released its comparison set. The one thing a designer actually wants to know — does the lettering come out right on a menu with eleven items — is still unanswered by the leaderboard.

What the API actually gives you

This is where the launch changes shape, and where most of the coverage stops.

The endpoint is POST https://api.canva.com/rest/v1/image-generations. You need a user access token with the asset:write scope and the image_generation capability, and you pin the model by passing craft_v1_2k in the model field. Jobs are asynchronous: you get an ID, poll for status, and receive a download URL that stays valid for 24 hours.

• Rate limit — 20 requests per minute per user.

• Prompt length — 1 to 10,000 characters.

• Idempotency — an idempotency_key is required, held for 24 hours, so a retry returns the original job rather than a duplicate.

• Billing — starting a job draws on the user's AI credit allowance; over the limit returns 429 credit_quota_exceeded.

• Failure modes — unsafe_input, timeout, model_not_available, model_retired, internal_error.

• Typical latency — Canva's own developer announcement puts a Craft job at roughly 50 to 60 seconds, which is consistent with a two-stage pipeline that plans before it draws.

Now the constraint that matters most. The API is in preview. Canva's reference states plainly that preview APIs may change without notice, that breaking changes will not get a new API version, and that public apps using preview APIs will not pass review and cannot be made available to all Canva users. That is the whole ballgame for anyone thinking about building a product on Craft: the model is generally available to Canva's users, but the route to it for third-party developers is not yet a stable contract.

The second constraint is aspect ratio. In Canva's editor the pitch is more than 30 aspect ratios plus custom sizes. Over the API there are six: square, landscape, portrait, landscape_3_2, portrait_2_3, portrait_4_5. There are no width or height parameters in the request; dimensions are decided by the model and reported back in the response. If your pipeline needs a 1,200 by 628 Open Graph card, you are cropping after the fact.

Third: pinning is a reliability decision. Canva recommends omitting model entirely so it can fall back to another model when one is unavailable — and states that if a pinned model cannot serve the request, the job fails rather than substituting. Pinning a model that launched yesterday means model_not_available is a branch your code has to handle on day one.

Canva's own docs send text prompts to a different model

The most surprising line in the whole release is not in the press materials. In the model parameter reference, Canva describes all three available models and what each is best for:

• craft_v1_2k — recommended for "infographics, posters, slides, people and portraits, abstract art, patterns, landscapes, nature, products, and marketing materials."

• lucid_origin — recommended for "architecture, logos, posters, or sci-fi and fantasy scenes."

• z_image_turbo — recommended "if the image includes text, or for food, drink, and general-purpose imagery."

Read that last line against the launch coverage. The pitch for Craft is a planning stage that decides text content and placement before rendering, and in-image text in 55 languages. The API reference, written by the same company on the same day, tells a developer who needs text in the image to reach for Z Image Turbo. Canva does include infographics and posters — both text-bearing formats — in Craft's own list, so this is not a flat contradiction. It is a hedge. And Canva labels the strengths as "based on Canva's evaluations," with an explicit instruction to test your own prompts before pinning a model. That is unusually candid for launch documentation, and it is the single most useful sentence in it.

The fair reading: the planner improves layout and makes targeted edits practical, but "legible text in 55 languages" is a model capability, not a guarantee that Craft is the best of Canva's three models at typography. Anyone whose use case is menus, posters or slide copy should test both Craft and Z Image Turbo on their own prompts rather than take either the launch post or this paragraph as the answer.

Why Canva built this instead of renting one

Canva did not build an image model because image models are fun. It built one because it was paying for other people's.

The context reported around the launch is unusually blunt. Canva was on track for roughly $4 billion in revenue, then cut its 2026 growth forecast, with reporting putting the revision from about 30% to about 20% and describing a valuation marked down from around $42 billion to around $35 billion. Coverage ties the pressure directly to AI inference costs — Canva had to pause part of its AI rollout in August 2026 because frontier-model token costs were not sustainable at its usage levels.

Owning the model changes that arithmetic in three ways: the marginal cost per generation becomes Canva's own compute rather than a vendor's price list, latency and quality stop being someone else's roadmap, and the model can be tuned on the asset Canva uniquely owns — a decade of template designs. The Leonardo.ai acquisition now reads less like a talent purchase and more like the seed of an in-house model lab.

It also changes Canva's competitive position in a way that cuts both directions. Craft is a design model, not a general image model: it is trained on layouts rather than photographs, and its selling points are composition and editability rather than photorealism. That makes it very good at the thing Canva sells and less interesting for the thing people use Midjourney for. The planned MCP integration with ChatGPT and Gemini is the tell — Canva wants to be the design layer inside other assistants, not to win a benchmark war against them.

What Craft costs, which is not a number

There is no per-image price, and Canva has not said which tier of its AI allowance Craft draws from.

• Canva Free — up to 20 combined Standard or Premium AI uses per month, so at most 20 Craft images.

• Canva Pro — $144 per year, up to 200 Premium uses per month, plus a separate small Ultra allowance.

• Canva Business — $250 per year per person, up to 400 Premium uses per month.

If Craft counts as a Premium tool, that works out to roughly 6 cents per image on Pro and about 5 cents on Business. If it counts as a Standard tool, paid plans draw no allowance at all within fair use. Canva's documentation is ambiguous on which applies, and the difference between those two outcomes is the difference between a metered product and a free one. For API callers the same ambiguity applies: jobs consume the end user's credits, so your cost is your customer's plan, not a line item you can forecast.

Complicated prompts consume more, so treat every per-image figure above as a ceiling rather than a rate. Free allowances reset at midnight UTC on the first of the month; paid plans reset on the billing date.

How to test Craft without betting anything on it

For Canva users, the decision is easy: open Dream Lab, generate a menu, a poster and a social card, and look at the small type. The model is free to try on every tier, and the planner's benefit should be most visible on layouts with several elements and a headline. Judge it on the format you actually produce, not on a landscape photo prompt — this model was not trained for that.

For developers, the calculus is different, and it starts with the preview status. Do not put a customer-facing product on a preview API that Canva says cannot pass app review. Do use it for internal tooling and prototypes, and write the integration the way you would write against an unstable contract: treat model_not_available as an expected branch, honour the 24-hour download window rather than storing URLs, reuse the idempotency key on retries, and do not hardcode the six aspect ratios into your own API surface — they are the most likely thing to widen.

If what you want is a comparison rather than an integration, the useful test is not "is Craft good" but "is Craft better than the three models above it, on my prompts." That test is cheap on the other side of the fence. Canva Craft is not on OrcaRouter — we do not serve it, so there is nothing here to route it through. The three models ranked above it are a different matter: gpt-image-2.5-sunburst, gpt-image-2.5-flare and gpt-image-2 are all callable through one API for more than 200 models with provider list prices passed through at 0% markup, so the same prompt can be run against all three on a single key and compared with whatever Craft produces in Dream Lab. The routing DSL lets you compose those calls into one request instead of three scripts, and automatic failover covers the models we do serve — it is not a safety net for a Craft integration, and you should not treat it as one.

That comparison is the decision most teams actually face this month: not whether to migrate to Craft, but whether the layout-planned output is worth a preview-only dependency and a 50-second job. One afternoon of the same prompt through both paths answers it better than a 1,341 Elo score does.

What to watch

Four things would change this picture, and all four are dated by Canva's own roadmap rather than by speculation.

• Craft-fast and Craft-ultrafast. The 50-to-60-second job time is the main practical objection to Craft in any interactive pipeline. If the fast variants land near a few seconds, the calculus changes for API users overnight — and Canva has not said whether they will carry the same model id or a new one.

• The preview label coming off. Until the generation API leaves preview, no public Canva app can ship on it. That single sentence is a bigger constraint on adoption than any benchmark.

• The MCP server. Shipping Craft into ChatGPT and Gemini would make Canva a supplier to the assistants it currently competes with, and would make the model reachable by a much larger set of developers than Canva's own API does.

• Whether the ranking holds. A 1,341 debut with a 65-point gap to first is a strong start for a first model, but the board moved 20 points in the third slot within twenty minutes of observation. What matters is where Craft sits once votes accumulate and independent comparisons exist.

The honest summary of day one: Canva has shipped a real model, with a real architectural idea behind it, into a developer surface that is still a preview and a pricing scheme that has not been priced. The part that is settled is the quality relative to its peers — fourth is genuinely good for a first attempt. The part that is not settled is everything a team would need to build on it.