AI Pricing Is Broken. Here's How Companies Are Fixing It

There's a number that keeps appearing in conversations about AI companies, and it's not what you'd expect. It's not the billions in funding or the exponential user growth. It's three.

Hypergrowth AI companies—the ones posting 100%+ year-over-year growth—change their pricing models at least three times every two years, according to data from Stripe. Meanwhile, low-growth companies? Only 22% make similar changes. The correlation isn't subtle.

Mayank Pant, billing solution architect at Stripe, spent a recent talk breaking down why AI pricing has become a competitive battleground, and why the old SaaS playbook is actively destroying margins for companies trying to use it. His central argument: the first price you set is a hypothesis, not a commitment. And the companies that treat it as such are the ones scaling.

Why SaaS Pricing Breaks in AI

Traditional SaaS companies enjoyed 80-85% gross margins that stayed relatively stable. You charged per seat, maybe added some usage tiers, and margins were predictable enough that you could set pricing once and revisit it annually.

AI economics don't work that way. Pant outlined three core problems: power users can consume 80% of your compute while representing 5-10% of customers. Infrastructure costs fluctuate unpredictably based on GPU availability and demand. And the feature you're charging premium prices for today might be commoditized industry-wide in six months.

"What might be a premium feature today in six months maybe a standard feature across," Pant explained. "So are you able to keep up with that pricing?"

The pace compounds the problem. The top 100 AI companies hit $20 million ARR in 20 months, versus 65 months for the top 100 SaaS companies—three times faster. When your product velocity is that high, static pricing becomes an anchor.

Stripe's data shows 84% of AI companies report that their product development outpaces their pricing updates. The gap between what they're offering and what they're charging creates either margin erosion (if you're underpricing new value) or customer frustration (if you're overcharging for commoditized features).

The Hybrid Model Migration

The industry response has been stark. In 2024, only 6% of AI companies used hybrid pricing—combining base subscription fees with usage-based charges. Now it's 41%, a seven-fold increase. Meanwhile, traditional seat-based and pure subscription models are declining.

Pant's explanation for the shift is straightforward: pure subscription leaves you vulnerable to power users burning through your margins. Pure usage-based makes customers hesitant to experiment because they don't know what their bill will be.

Hybrid pricing splits the difference. A base fee establishes customer commitment and provides revenue predictability. Usage fees on top let customers scale their use while protecting company margins. "You are not alienating any category of the users," Pant noted.

But implementing hybrid pricing well requires rethinking how you communicate value. Pant walked through a five-step framework that several high-growth AI companies are using.

Defining Value in Customer Terms

The first step is the hardest: defining value from the customer's perspective, not yours. Pant used Gamma, an AI presentation tool, as his example. As an engineer, you might think in terms of API calls and tokens consumed. As a customer, he doesn't care about any of that. He cares about how many quality presentations he gets.

"For the customer it is the quality of the deck and the relevancy of that deck," he said. The technical infrastructure is invisible—and should stay that way.

Pant broke AI value propositions into four categories. Automation saves time and therefore cost. Augmentation keeps headcount stable but increases output quality. Enhanced service provides access to proprietary data or capabilities. Improved results directly impact business metrics—like Intercom pricing based on support tickets resolved without human intervention.

Once you've identified which value category you're in, you can define your charge metric—the billable unit that represents that value. This is where most companies make their first mistake: they default to technical metrics (tokens, API calls) because those align with their costs. But customers don't think in those terms.

Pant suggested three approaches. Consumption-based (API calls, compute time) is easiest to implement but hardest to align to value. Workflow-based (images generated, documents summarized) aligns to the product. Outcome-based (candidates hired, leads qualified) aligns to customer ROI but is hardest to attribute.

The emerging pattern is to abstract technical metrics behind credits. You sell customers 100 credits per month. Under the hood, you're constantly adjusting what those credits buy—five API calls of one type, ten image generations, whatever. The customer sees a stable number. You maintain pricing flexibility.

Guardrails Against Bill Shock

Flexible pricing creates a trust problem. If customers don't know what they'll be charged, one unexpectedly high bill can undo months of goodwill. Pant emphasized building guardrails from the start: usage caps that stop service or require top-ups, automated notifications at 50%, 70%, and 90% of limits, and rate limiting to prevent runaway costs from bugs or misuse.

"A wrong bill can erode a lot of customer trust," he said. "You might be doing great for three, four, five months, but if on the sixth month your bill goes wrong, bill goes very high, then those customers go."

The design principle is simple: fair pricing, no surprises. Customers should always feel in control of their usage and costs.

The Infrastructure Question

All of this requires billing infrastructure that can actually support rapid iteration. If changing your pricing model requires three months of engineering work, you can't iterate fast enough to stay competitive.

During Q&A, an attendee asked about the risk of frequent pricing changes frustrating customers. Pant's answer revealed the sleight of hand that makes this work: customer-facing prices can stay relatively stable while you adjust what features map to which tiers under the hood.

Grandfathering helps too. Existing customers keep their pricing when you launch new versions. New customers pay more. The credit abstraction layer gives you room to maneuver without constantly renegotiating with your user base.

Another question probed how major AI labs use iterative pricing despite appearing to have static tier structures. Pant's response: "They use it under the hood." The four-tier structure (good, better, best, enterprise) stays constant. Features migrate between tiers. The credit system abstracts the changes.

What's striking about Pant's framework isn't that it's revolutionary—it's that it codifies what hypergrowth companies have already figured out through trial and expensive error. The 78% of AI companies building on Stripe includes Anthropic, OpenAI, ElevenLabs, and Intercom. They're all using hybrid models. They're all iterating frequently.

The companies still trying to force AI products into SaaS pricing templates are the ones struggling with margins, customer acquisition costs, or both. The infrastructure costs of AI aren't going down fast enough to make traditional models work. And the product velocity isn't slowing.

Which means the only variable left to optimize is pricing itself—and the speed at which you can change it.

Marcus Chen-Ramirez is senior technology correspondent at Buzzrag.