Canva cuts its growth forecast from 30% to 20% — and launches its own image model, Craft

Canva cuts its growth forecast from 30% to 20% — and launches its own image model, Craft

Canva launched its first proprietary image generation model, Canva Craft, on 22 October 2026 — barely three months after the company cut its 2026 growth forecast from 30% to 20%, precisely because serving other companies' AI models had become too expensive. With Craft, Canva is moving inference in-house, and the model debuted straight at fourth place on DesignArena's crowdsourced leaderboard, above models from Google, Microsoft, Meta and xAI.

What Craft is, and how it works

Craft differs from generic image models by being built design-first. According to Canva's announcement, as reported by The Verge, the model uses a two-stage architecture that plans the design before the image is generated — which the company claims delivers "greater precision, consistency and control". That is a company statement, not an independently verified property, but the architectural idea is concrete: rather than producing an image in one straight path from text to pixels, the model first reasons about the composition.

The specifications are equally concrete. Craft generates images at up to 2K resolution and in more than 30 aspect ratios, and the model was trained on millions of human-made templates from Canva's own content library. That last point is Canva's clearest competitive advantage: the company holds a corpus of design material that almost no one else has.

The training process, as TechTimes describes Canva's own account, followed pre-training with supervised fine-tuning and reinforcement learning in which Canva's professional designers assessed 25,000 head-to-head comparisons to calibrate the model's judgement. The version in production is identified internally as CRAFT-1 2K, according to TechTimes — a single-source detail not confirmed by other outlets.

Where Craft actually ranks — and what Canva claims

Here it is essential to draw a sharp line between benchmark data and marketing.

The crowdsourced DesignArena leaderboard, where users compare AI-generated designs blindly, places Craft fourth at launch, behind OpenAI's GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2, The Verge reports. TechTimes gives an ELO score of 1,341 at its debut on 22 October — above models from Google, Microsoft, Meta and xAI. This is a benchmark position, not an objective measurement of "best", and the leaderboard rests on crowdsourced judgements with all the limitations that entails.

Canva nonetheless made a far stronger claim. On its official X account, quoted by Digital Trends, the company wrote (translated from English): "With Craft, Canva is now the world's second most advanced AI lab for image generation on @DesignArena's global leaderboard."

The claim does not literally match a fourth-place ranking behind three OpenAI models. The most obvious explanation is that Canva aggregates OpenAI's models into a single actor — behind "OpenAI" as a lab, but above all other companies. AFR reports the same framing: an image tool that, according to Canva, ranks higher than Google's and is surpassed only by OpenAI. It remains unresolved whether that aggregation is justified. There is no independent technical evaluation of Craft yet, and the claims of "greater precision, consistency and control" rest on Canva's word.

Why the launch comes now: the costs

What makes Craft more than a product story is the business logic behind it. In August 2026, CEO Melanie Perkins told investors that Canva was cutting its 2026 growth forecast from 30% to 20% — a one-third reduction — because the cost of serving a single AI task had become too high, according to TechTimes. The AI features, in other words, cost more to deliver than the business model could bear, and according to AFR, the costs around AI token usage led the company to scrap planned, promoted AI launches.

The response has two legs. One is Craft itself: by owning and operating the model, Canva can control inference costs instead of paying per call to third-party providers. The other is the roadmap. Canva is already developing two faster, more cost-effective variants, Craftfast and Craft-ultrafast, Digital Trends reports — a clear signal that the goal is cheap volume, not just top quality.

The company's valuations have followed the economics downward, but the sources diverge. TechTimes writes that investors Blackbird Ventures and Airtree revised the valuation down from roughly $42 billion to $34.9 billion in August. AFR reports that the largest investors cut $10 billion from valuations that, according to the newspaper, "now sit at $50 billion". The figures cannot be reconciled from the available coverage, and neither can be presented as the correct one. What both share, however, is that Canva has taken a market-value hit in the wake of its AI costs.

Who can use it, and what's coming

Craft is available to all Canva users through the Dream Lab tab, located in the panel at the bottom left of the interface, according to Analytics Insight. Developers can use the model through Canva's public API.

The roadmap points beyond image generation. According to Digital Trends, Craft's outputs can be converted into layered, editable designs via Magic Layers — that is, from flat pixel graphics into something that can actually be edited in Canva, which is the point for the design tool's users. Canva also plans integrations with third-party AI assistants via its MCP implementation. The Craftfast variants are, per the company's own communication as reported by Digital Trends, intended to make generation both faster and cheaper.

What remains to be answered

Several questions remain open. Whether Craft's quality holds up under sustained use depends on independent evaluations, which are so far lacking. DesignArena's crowdsourced methodology measures preferences in short comparisons, not performance in actual design workflows. The "second most advanced AI lab" claim is Canva's own framing, in conflict with a literal fourth-place ranking, and unresolved. And the core cost thesis — that in-house models will reduce Canva's AI spending enough to reverse a ten-percentage-point growth downgrade — can only be assessed once the Craftfast variants and API usage have been running for a while.

What is documented, however, is the pivot itself: a company squeezed by AI inference costs into revising its financial outlook is responding by becoming an AI player in its own right — with an early benchmark position that makes the claim less speculative than most "we're building our own model" announcements.