Canva introduced Craft on October 22, 2026, as its first image foundation model trained from scratch. It generates images up to 2K resolution and is available in Canva and through a public API preview. Its design-focused architecture, editable-output workflow, and developer access each have different practical boundaries.
The short answer: Craft can generate images now; Magic Layers provides a separate path to editable designs; the image-generation API remains a preview, not a generally available production contract.
What arrived on October 22—and what did not
Canva’s developer announcement identifies the model as craft_v1_2k, with output up to 2K. The image-generation endpoints themselves entered preview on October 1, 2026, according to Canva’s changelog. October 22 therefore marks Craft’s addition, not the invention or general availability of that API.
The launch materials do not announce downloadable Craft weights or a self-hosting license. Hosted API access should not be described as an open-weight release. Craft-fast and Craft-ultrafast are upcoming variants in the announcement, not releases established by it.
Craft-specific MCP support is also marked forthcoming in the October 22 developer post. That is distinct from Canva’s existing MCP server, whose broader design capabilities are already documented. An available connector does not establish that a particular model is exposed through it.
Why the editable-design handoff matters
Craft separates composition planning from rendering. A planning stage specifies elements, relationships, scale, hierarchy, and text before a rendering stage produces the image. Canva says this structure supports Magic Layers conversion and more targeted control.
Magic Layers is not simply another name for Craft. Canva’s October 21 research overview describes a learned decomposition model that interprets a composition and produces editable text and elements. It also explains the company’s emphasis on brand consistency, cultural context, and controllability—not merely attractive pixels.
Consider a hypothetical campaign graphic for a ceramic bottle. The useful acceptance test is not whether the initial image looks impressive. It is whether a marketer can change the offer, move the bottle, and correct the headline without rebuilding the composition. That is our operational interpretation of the handoff, not a reported firsthand test.
For a pilot, keep the original generated asset and the edited design as separate review artifacts. Ask a reviewer to verify lettering, object boundaries, and layout after conversion. Measure time to an approved revision rather than celebrating the first generated draft.
The preview API changes the automation decision
Canva’s current generation documentation warns that preview APIs can change without notice and without a new API version. Public apps using them cannot pass review for availability to all Canva users. This is a distribution constraint, not just a caution about occasional bugs.
The API generates an image, not automatically a layered Canva document. Jobs use the connected user’s AI allowance; results can be saved as image assets, while download links expire after 24 hours. Developers should not equate successful image delivery with a completed editable-design workflow.
There is also a format distinction. The launch advertises more than 30 aspect ratios and custom ratios, whereas the API reference currently enumerates six presets. Treat the endpoint’s schema as the integration contract rather than assuming the product interface’s flexibility transfers unchanged.
Our recommendation is a bounded internal experiment with explicit human approval, not a customer-facing rollout dependent on preview app approval. Keep generation, editable conversion, brand review, and publication as separate workflow states. That makes failures visible instead of letting a plausible-looking image silently become a published campaign.
What business teams should watch next
Canva explicitly identifies fine print and technical drawings as areas for improvement, with advanced editing and reference-image capabilities slated for future releases. Those statements are reasons to validate demanding assets, not evidence that the limitations have disappeared.
The next meaningful milestones are a changed API preview designation, documented Craft-specific MCP access, and evidence that editable revisions survive real campaign requirements. Until then, evaluate Craft as a newly available image model with an editing pathway—not an autonomous, unrestricted design-production system.