Last month, developers using Google’s Antigravity AI tool ran into a weird problem: they couldn’t tell what they were paying for. Google introduced a new pricing model based on credits ($25 buys 2,500 of them), but stopped short of explaining what a credit actually represents.
At the same time, users reported that their usage limits were shifting all over the place. Workflows that once ran freely were suddenly hitting caps. The same tasks seemed to consume wildly different amounts of quota.
It was like going into a grocery store and getting wildly different prices for the same basic product. The cash register acted like a slot machine.
One day, a gallon of milk costs three dollars. The next day, it costs seven because more shoppers walked into the store, the refrigeration units were working harder, or the cashier decided your basket looked “computationally expensive.” You pull the lever and hope the total lands somewhere reasonable.
That uncertainty changes customer behavior. People stop experimenting. They hesitate before clicking “run.” They start treating every interaction like a gamble instead of a service. When customers can’t predict the cost of an action, they begin optimizing for fear rather than value.
And unlike a casino, where the volatility is part of the entertainment, enterprise software depends on trust and predictability. CFOs cannot budget against “maybe.” Procurement teams cannot negotiate around “it depends.” Developers cannot architect products when the underlying economics shift from one API call to the next.
Every AI interaction requires compute, but that consumption can vary dramatically depending on the task, the model, and the context window involved. Providers therefore wrap all that complexity into a single abstraction: the credit. Roughly 30% of new AI services have now adopted credit-based models as a monetization mechanism.
From the provider’s perspective, this makes perfect sense. Credits act as a buffer against volatile infrastructure costs. But from the customer’s perspective, it can feel like providers are simply passing infrastructure volatility directly onto their shoulders.
To be fair, vendors are trying to solve a real problem here. Credits are not inherently bad. They are a solution to a specific problem. They’re just not the solution to every problem. If you don’t have to use them, don’t. Long live Occam’s Razor.
There is no “holy grail” pricing model for AI. No single pricing construct is going to work across every customer or workflow. In reality, most software companies will end up operating multiple pricing models simultaneously. You may have subscriptions for core platform access, usage pricing for AI-intensive workloads, and credits layered on top for highly variable consumption patterns.
But right now, too many companies are simply passing infrastructure volatility through to customers and calling it pricing. And as my colleague Michael Mansard notes in his excellent white paper on AI credits, “a response is not a strategy.”
The winners in AI monetization will be the companies that best translate infrastructure complexity into pricing models customers can actually understand and trust.
Because no one wants their grocery store to act like a casino.
