Canva's First Image Model Attacks Adobe's Seat Pricing From Below — and the Real Story Is Cost, Not Quality

• Two stages: a planning model interprets the prompt and writes a composition plan, then a rendering model turns that plan into the image.
• Up to 2K resolution across more than 30 aspect ratios, text in 55 languages, and prompts accepted in any language.
• Trained on millions of human-made Canva templates rather than scraped photographs, so output inherits Canva's layout and typography logic.
• Post-training used supervised fine-tuning plus reinforcement learning with feedback from Canva's professional designers, who judged more than 25,000 head-to-head pairs.
• Distributed at both ends: in Dream Lab for Canva users worldwide, and to developers through the Connect API as model craftv12k, where each job draws down the caller's AI credit allowance.
• Two cheaper variants, Craft-fast and Craft-ultrafast, are promised at a fraction of the cost of leading frontier models, and a Canva MCP server will let assistants such as ChatGPT and Gemini generate directly.

Historically, the cost to create a design cost us essentially nothing. The cost of goods to serve a user has gone from cents per month to many, many more cents.

• Craft beat Alphabet's, Meta Platforms's, ByteDance's and xAI's image models in independent testing, and Canva now serves it through an API and an MCP server — erodes the frontier labs' lead in design-grade images.
• Canva calls itself the second-most-advanced image-generation lab in the world. That is a company claim, but it is a fast climb from a standing start.
• Craft's output converts into editable layers inside Canva through Magic, so a generated image keeps feeding the same workflow a seat licenses.
• Adobe holds the counterweight: more than 1 billion monthly active users, creative freemium MAU above 100 million growing more than 70% year over year, and Firefly ending ARR up 40% quarter over quarter.

• Figma doubled the credits every Professional subscription buys at the same price on Aug. 25, 2026, and cut Figma Make credit prices by roughly 50% in September — buys volume with margin.
• Figma's agent and Weave moved to general availability on Oct. 6, 2026 and now consume AI credits, so usage inside the product is billed by usage.
• Wells Fargo cut Adobe to Underweight with a $225 target from $270 on Oct. 12, 2026, citing an enterprise CIO survey in which Adobe lost budget share, a difficult Q4 bookings ramp, and the absence of a price increase.
• One bullish counterpoint on Adobe: the shares trade at roughly 8.7x forward adjusted earnings, with a trailing free-cash-flow yield near 10%, so much of the disruption may already be in the price.

We recently reduced the price of Figma Make credits by around 50% in some cases, prioritising volume over near-term margins, because we think the prices were too high relative to perceived value.

• NVIDIA's DGX Cloud is available through AWS Marketplace Private Offers for training and customising models, and Leonardo.ai — part of the Canva family — is named as an early adopter running it on AWS.
• NVIDIA's revenue reached $96.22B in the quarter ended July 26, 2026, up 106% year over year, at a 75% gross margin. Fragmenting the model layer has not slowed the silicon cycle.
• Amazon spent $54.2B on property and equipment in the June 2026 quarter against $45.4B of operating cash flow, funding capacity ahead of the demand curve.
• Alphabet spent $44.9B on capital expenditure in the June quarter on $119.8B of revenue, and Meta Platforms spent $30.1B on $60.8B — the same bet at different intensities.
• Downstream, Canva's MCP server aims to put Craft inside assistants such as ChatGPT and Gemini, and its Connect API lets developers call craftv12k directly — Canva starts selling generation to the labs that were its suppliers.

• Adobe deferred a core creative price increase while its subscription cost line grew 24%, and it cannot raise seat prices into a market where one free prompt produces a comparable asset — puts Adobe's seat pricing under a deadline it cannot defer.
• The thesis fails if the frontier labs cut prices faster than platform labs can cut inference cost: OpenAI's models sit above Craft on the leaderboard and can defend with a price cut rather than a model release.
• Canva's 20% guidance assumes the 90% cost reduction holds as usage rises. If inference costs re-accelerate, the same bill that cut growth once will cut it again.
• If Craft moves from fifth into the top three in marketing materials, the price argument becomes a product argument and the timeline compresses sharply.