The price of intelligence fell in July. The price of the intelligence Canva was buying did not.
In late July 2026 OpenAI repriced the GPT-5.6 family. Luna, the fastest and cheapest tier, went from $1.00 to $0.20 per million input tokens and from $6.00 to $1.20 per million output tokens, a cut of 80 percent on both sides. Terra, the middle tier, went from $2.50 and $15.00 to $2.00 and $12.00, a cut of 20 percent. Sol, the flagship, did not move. It still lists at $5.00 and $30.00 (Forbes, July 31, 2026; CloudZero OpenAI pricing, August 2026).
The headline was that AI got 80 percent cheaper. Run the ratio instead. Before the cut, Sol cost five times Luna on input and five times on output. After it, Sol costs exactly twenty-five times Luna on both. The repricing did not narrow the distance between the top and the bottom of the ladder. It multiplied that distance by five.
Two software companies published what that distance costs
Canva cut its 2026 revenue growth target from 30 percent to 20 percent, a third of the target, in its second-quarter update to shareholders. Demand was not the problem. Canva reported second-quarter 2026 revenue of $921.9 million, up 25.2 percent year over year, against an annualised run rate near $3.69 billion (Startup Daily, August 5, 2026). Cofounder Melanie Perkins wrote that the "average cost of serving an AI task was too high" and that the company had been "relying too heavily on frontier models".
Perkins named the failure before she named the fix. "Several of our first-party models were not yet ready for release, and our pricing, consumption model and usage controls had not caught up with the outsized demand we were seeing." Three of the four things she lists are not models. They are meters. Canva has since cut cost per AI task by close to 90 percent since Canva AI 2.0 shipped in April 2026, and Perkins framed the slowdown as deliberate: "Rather than broadly rolling out a product before the underlying economics were ready, we decided to slow the rollout while we rebuilt the architecture, reduced unit costs and strengthened the business model" (Fortune, August 12, 2026).
Wix ran the same play earlier and said so on its second-quarter call on August 4, 2026. CFO Lior Shemesh put it in one line: "A company-wide priority for 2026 was to lower inference costs with our own LLMs." Wix shipped Base1, a model trained in house, in June. CEO Avishai Abrahami said its results on the task beat the frontier providers Wix had been using, and that "it also costs dramatically less". Shemesh guided Base44 non-GAAP gross margin to approximately 60 percent in the second half of 2026, "a very significant improvement from the near 0 gross margin entering 2026" (Wix, Q2 2026 earnings call).
That 60 percent is guidance, not a reported result, and it deserves to be read as one. Wix's consolidated non-GAAP gross margin in the second quarter was 67 percent, down three points year over year, which Shemesh attributed in part to elevated AI compute costs. The in-house model has not yet reached the consolidated line. Canva is private, and its figures are self-reported and unaudited. Both companies serve consumer-scale volume, which is the condition that makes training your own model pay. Neither case is a template for a bank.
The tier that got cheaper is not the tier that broke the margin. Hikari Blue · operator note
What this changes for a firm that will never train a model
Most regulated enterprises will not train a frontier-class model, and should not plan to. The transferable finding is not build your own. It is that cost per task is a variable under management control, and that both companies moved the same two levers: they pushed work down the tier ladder, and they metered it. Metering exists at any volume. Perkins listed pricing, consumption model and usage controls as the things that had not caught up. None of those require a training run.
The July repricing changed the arithmetic of that choice, and not in the direction most readers assumed. Route a task from the flagship down to the cheap tier and what survives is the cheap tier's price. In June, that residue was 20 percent of the flagship cost. Today it is 4 percent. The saving per token barely moved in dollars. What moved is the floor: the cost that remains after a routing decision is now five times smaller than it was two months ago. That is the whole case for treating model routing as a governed decision rather than a default in a config file.
The constraint that stops firms capturing this is measurement, not architecture. A routing policy needs a cost per task attributable to a feature and to a customer, available before launch, not a cloud invoice reconciled at month end. Canva learned its unit economics from its users. Wix set the target before the margin broke. Both statements sit in the public record. Only one of them cost a third of a growth forecast.
What to examine before the next AI feature ships
Four questions decide whether an AI feature earns its place. What does one completed task cost, at the tier it actually runs on, not the tier it was prototyped on. Which tasks are on the flagship because they need it, and which are there because nobody revisited the default. What usage control caps a single customer's consumption, and who is paged when it trips. What price is charged for the feature, and does that price survive the cost per task at the ninetieth percentile of usage rather than the mean.
The metric to track is cost per completed task, by feature, held next to the revenue that feature carries. Not monthly AI spend. A firm reporting monthly AI spend is reporting a bill. A firm reporting cost per completed task is reporting a margin, and only the second one can be defended in a board meeting. This is the ordinary work of an AI operating layer: the layer that decides which model answers, records why, and prices the answer.
What does one completed AI task cost you today, and who checked before you shipped it?
Model prices will keep falling on the tiers with competition. The flagship held its price through the largest cut of the year. A cost line that only improves when a vendor decides to improve it is not a cost line you manage.
- Rachel Wells (July 31, 2026). OpenAI Cuts GPT-5.6 Pricing Up To 80%, As AI Costs Come Under Scrutiny. Forbes. Source for the July 2026 repricing of the GPT-5.6 family: Luna at $0.20 input and $1.20 output per million tokens after an 80 percent cut, Terra at $2.00 and $12.00 after a 20 percent cut, and the flagship Sol receiving no price cut. forbes.com, OpenAI cuts GPT-5.6 pricing up to 80%
- CloudZero (August 2026). OpenAI API pricing in 2026: every model after the July price cuts. Source for the post-cut price table used for the ratio arithmetic: Sol unchanged at $5.00 input and $30.00 output per million tokens, Terra at $2.00 and $12.00 down from $2.50 and $15.00, Luna at $0.20 and $1.20 down from $1.00 and $6.00. cloudzero.com/blog/openai-pricing
- Mia Osmonbekov (August 12, 2026). Canva, the rare startup that grew fast and made money, sees AI cut its growth forecast by a third. Fortune. Source for the revised 2026 growth forecast of 20 percent, the reduction of cost per task by nearly 90 percent since Canva AI 2.0 launched in April 2026, and the Melanie Perkins quotation on slowing the rollout to rebuild the architecture and reduce unit costs. fortune.com, Canva growth forecast cut by a third
- Startup Daily (August 5, 2026). Canva cuts revenue forecast by a third as it tackles high AI costs. Source for second-quarter 2026 revenue of US$921.9 million, growth of 25.2 percent year over year, the annualised run rate of approximately US$3.69 billion, the move of the full-year growth forecast from 30 percent to 20 percent, and the Perkins quotations on the average cost of serving an AI task, on relying too heavily on frontier models, and on first-party models, pricing, consumption model and usage controls. Reporting on the Q2 CY2026 shareholder update. startupdaily.net, Canva cuts revenue forecast by a third
- Wix.com Ltd. (August 4, 2026). Q2 2026 earnings call transcript, The Motley Fool. Source for Lior Shemesh on lowering inference costs with the company's own LLMs as a company-wide priority for 2026, on Base44 non-GAAP gross margin guided to approximately 60 percent in the second half against near zero entering 2026, on consolidated non-GAAP gross margin of 67 percent in the quarter and elevated AI compute costs, and for Avishai Abrahami on Base1 results and cost. fool.com, Wix Q2 2026 earnings call transcript
The Hikari Blue team · Austin, August 2026