Corazza says Canva models cost 40 times less for design

The short version

Stefano Corazza says Canva’s smaller models cost 40 times less than the most expensive frontier models on design tasks.

Canva’s proprietary data and control over serving help, but Corazza gave no method for the comparison.

Compare models on the cost of an approved result, including revisions and human review.

---

Canva’s AI research head Stefano Corazza put the gap at 40 times on The TWIML AI Podcast 6:14. He said its smaller models can do the same design task for less than the priciest frontier models.

Canva sells design software and has a stake in making its own AI cheaper to run. But Corazza did not name the models, task or measurement behind that comparison.

If you buy or build AI features, the useful question is narrower than whether small models win. By the end, you’ll have a way to compare costs against the quality your team actually needs.

Canva trains where its data gives it an edge

Corazza named two reasons for building models. Canva has design data other companies lack, and it can integrate models into its product. He said that combination can make a smaller model competitive on a specific task.

The company did not take one route for every model. It trained Craft, its image model, from scratch. For Brush, its design creation and editing model, Canva started with an open-weights model and built on it 9:52.

So that distinction matters for a smaller company. Training from scratch demands a different commitment than adapting an existing model.

Neither route is justified by a low serving cost alone. The comparison also needs to include the work of collecting data, testing quality and maintaining the system.

Danila Shtan sees a related pattern in a different customer base. He said most of his customers modify open models for narrower work. That supports testing specialization, not assuming Canva’s savings will transfer to another workflow.

Editable layouts change what a model must produce

Canva’s product design also changes the model’s job. Craft expands an input, turns it into a structured scene and then generates pixels. The scene description includes positions and bounding boxes, which define where elements sit.

So a person can intervene before the final image generation. Corazza described a designer moving an element rather than repeatedly asking an image model to interpret a small positional change.

“And so it’s a lot easier for a designer to just move that element to the right and then regenerate.”

Stefano Corazza, on editing a design 13:37

The point is not that every element can change without regeneration. Corazza described moving elements and generating pixels afterward. Canva also offers Magic Layers, which separates a flat image into editable elements, including images made outside Canva.

Still, a product team can apply the same question without building a model. Which parts of an output must remain editable after generation? Preserving those parts may reduce repeated requests, although Canva did not give a saving for that step.

Corazza gave no method for the 40 times comparison

Canva has given other cost comparisons, but they do not verify Corazza’s figure. Startup Daily reported in August that Canva’s image model was 30 times cheaper than a frontier model. The report gave a 17 times comparison for video.

• The comparison model remains unnamed in Corazza’s 40 times figure.
• The design task and the quality threshold are not specified.
• The cost boundary does not say which serving expenses are included.

Yet those are different comparisons across different kinds of work. The report does not identify a method that lets a reader reconcile them with Corazza’s design-task figure. None establishes what a customer would save on a finished design.

For instance, Canva says Brush costs 46% less per edit than its previous design agent. OrcaRouter’s review says that is Canva’s internal comparison. It also says Canva has not identified what powered the previous agent or published a sample size.

That does not make the number false. It makes the number hard to use in a budget. No independent study found in the supplied research compares Canva’s models with frontier models on design quality.

Design quality needs a test beyond first drafts

Corazza said many benchmarks focus on an initial generation rather than editing an existing design 23:21. Canva uses structured criteria for properties such as readability and layout. It also uses human evaluation.

The PresentBench study by Xin-Sheng Chen and colleagues shows how detailed a generation test can be. It includes 238 slide-generation instances and an average of 54.1 checklist items per instance. The authors report stronger alignment with human preferences than existing methods.

But PresentBench tests slide generation, not Canva’s design editing. The available abstract does not give a count of human evaluators. Its result cannot establish whether Brush edits a customer’s existing deck well.

For example, Corazza said Craft’s early cartoon images looked too realistic and too three-dimensional. Canva found the mismatch through arena testing and was working to tune the style. A strong result on portraits did not establish quality across cartoons.

Faster revisions can spend the savings again

Brush works with layered designs and uses tools to change them. Corazza said it can inspect what it generates and revise its work before showing a person. Faster inference makes more such iterations possible within a wait a user will accept.

“the faster you make that iterative loop, the more effective everything is”

Stefano Corazza, on Brush’s revision loop 35:18

But more iterations also mean more model calls. A cheaper call can produce a more expensive job if the system needs enough retries. The relevant bill includes the model, the tools it invokes and any human correction afterward.

Corazza did not give a per-design cost for a full revision loop. So teams cannot use the 40 times figure as that number. They can record how many attempts each candidate model needs and whether the final result passes review.

And that measurement matters outside design too. Woodson Martin said many enterprise tasks do not need frontier models. His point concerns enterprise workloads generally, not a measured result for Canva’s designs.

More control over serving brings more optimization work

Corazza said Canva optimizes inference across hardware, model steps and serving 41:11. Inference is the work of running a model to produce an output. Canva uses cloud hardware rather than owning the machines.

He said access to model weights and self-hosting give the team more room to optimize. But that control also requires people who can work across models, graphics and cloud systems. A company renting an API can still compare providers without taking on that operation.

The decision depends on volume as well as price. A saving on each call may not cover the staff and infrastructure needed to capture it. Corazza did not provide Canva’s complete development or operating costs, so the interview cannot settle that calculation.

Compare the cost of an approved design

If your team pays for AI generation or editing, pick a recurring task this week. Use existing work that reviewers can assess, including requests to revise an earlier output. Compare candidates on the same cases rather than their published price alone.

• Set approval criteria for the task, including the changes a reviewer must be able to make.
• Record every attempt and its cost, including tool calls and retries.
• Count review time and identify outputs that still need human repair.
• Compare accepted results by dividing total test spend by the number reviewers approve.

Keep human review time beside the cost calculation if you cannot price it reliably. And check whether one model fails on a category that matters to your customers. Canva’s cartoon example shows why an average quality score can hide a useful distinction.

Corazza said Canva planned faster versions of Craft in the coming weeks 48:17. Watch whether Canva publishes their prices and quality comparisons. A faster model is useful only if its results still meet the needs of the task.

Copy for your team. For this workflow, compare candidate models on approval rate, review time and cost per approved result.

Send this to the person measuring your AI feature costs.

Watch the moment Corazza gives Canva’s cost comparison. 6:14