Pricing AI Products When Every Query Costs You

Start with the one fact that breaks the old playbook

Price your AI product by working backward from what a single unit of value costs you to deliver, then choose a model that hands that variable cost back to the customer in a form they will accept. That is the whole job. Everything else is packaging.

The reason AI product pricing is hard, and why copying a SaaS pricing page will quietly bankrupt you, comes down to one fact: AI has a marginal cost that scales with usage. Classic software was close to free to serve. You paid to build it once, then the cost of the next customer rounded to zero, which is why margins ran at 80 to 90%. AI breaks that at the foundation. Every query, every inference, every agent action consumes real compute you pay for.

There is even a name for it now. Analysts call it the token tax: the structural levy any company pays for building on inference it does not own. And it is not small. At scaling-stage AI companies, inference alone eats roughly 23% of revenue, dragging gross margins 15 to 30 points below the old SaaS standard.

What the token tax does to your margin

Here is the math in its simplest form, from The SaaS CFO. Take a product doing $100 of revenue with $20 of traditional COGS, so an 80% gross margin. Bolt on AI features that add $15 of variable cost for inference, routing, and supporting infrastructure. Revenue stays $100. COGS climbs to $35. Your gross margin just fell from 80% to 65%, before you account for heavy users, sloppy prompts, or expensive model choices.

The aggregate numbers confirm it is structural, not a startup hiccup. Pure AI-first companies run at 50 to 60% gross margins today, against 80 to 90% for mature SaaS. ICONIQ's data shows the AI cohort improving, from 41% in 2024 to 45% in 2025 to 52% in 2026, but the ceiling sits well below SaaS norms. The fastest-growing startups are often worse off, running near 25% gross margins, sometimes negative.

The trap is that the cost is invisible at trial scale. It only shows up when usage compounds. Microsoft reportedly pulled back internal Claude Code usage during this year's cost scrutiny, and Uber burned through its entire 2026 AI budget in four months. If sophisticated buyers blow their budgets this badly, your pricing page is not protecting you by accident.

Why per-seat pricing is collapsing

Per-seat made sense when a seat's cost to serve was fixed and tiny. It is breaking because a seat's cost is now variable and unbounded. Charge a flat $80 per user and one power user running an agent all day can wipe out the margin on ten light users. The price no longer tracks the cost.

The market is voting. Per-seat pricing fell from 21% to 15% of SaaS in twelve months, while usage-based pricing has gone the other way, climbing from 30% adoption in 2019 to about 85% by 2024. The pendulum is swinging from "pay for access" to "pay for what you consume" because, on the cost side, that is now how the product actually behaves.

The five models, and when each one fits

There is no single right answer. The category drives the model, and most products end up blending two.

• Per-seat still works where AI is a light assist on top of human work and per-user consumption is predictable. Productivity tools lean here. It is simple to buy and easy to forecast. It just stops working the moment usage gets heavy or agentic.
• Usage-based (per token, per call, per generation) aligns price directly with cost. Developers and API products accept it readily. The downside is real friction: usage anxiety kills adoption because buyers hate unpredictable bills, and enterprise procurement wants a number it can budget.
• Prepaid credits are the underrated middle path. The customer buys a pack of credits upfront and spends them across features. It converts unpredictable usage into a predictable purchase, removes the running-meter anxiety of pure usage billing, and floats you the cash before you incur the cost. The catch is that you have to set the credit-to-cost ratio carefully and decide whether credits expire, or you end up carrying a liability of unspent balances.
• Outcome-based charges only when the AI delivers a defined result. Intercom's Fin charges about $0.99 per resolved conversation. It is the most value-aligned model and the hardest to operate. You have to define "success" precisely and measure it cleanly, or you will fight disputes over what counts. One enterprise contract reviewed by Deloitte pays $18,000 in a month only if the agent's resolution rate clears 80%.
• Hybrid is the answer most successful companies land on: a base subscription for predictability, plus a usage, credit, or outcome layer that scales revenue with consumption. It is now the default. Estimates of how widespread vary with how you count, from around 41% of products to over 90% of AI software companies using some hybrid element, but the direction is not in dispute. It protects your margin while keeping the entry price low enough for product-led growth.

If you sell to enterprise, lead with a subscription and put usage underneath, because procurement needs predictability. If you sell to developers, usage-based is fine on its own. If you sell to consumers, a flat subscription or a prepaid pack almost always beats a live meter, since consumers will not tolerate a counter running in their head.

The lever most teams miss: drive your own cost down first

Pricing is half the equation. The other half is making the unit cheaper to deliver, which widens whatever margin your pricing captures.

Two levers matter most right now. First, prompt caching: Anthropic and OpenAI both offer roughly 90% discounts on cached input tokens. If your product has a stable system prompt and repeated context, treating caching as a P&L decision rather than an engineering footnote can cut effective per-query cost by an order of magnitude in about a week of work. Second, model routing: send easy requests to a cheap model and reserve the expensive one for hard tasks.

The macro tailwind helps too. Inference prices fall fast, what a16z calls LLMflation, roughly a 10x annual decline for equivalent performance. But do not bank on it as guaranteed relief. Jevons paradox bites: as each call gets cheaper, teams deploy AI to more use cases and total token consumption rises faster than unit cost falls. Cheaper tokens without usage governance just means you spend the same money in more places.

A practical sequence for setting your price

If you are pricing an AI product from scratch, work in this order:

1. Instrument before you price. Measure the fully loaded cost of one typical task, including the tokens the user never sees, retries, and infra. Most teams have never modeled what one unit actually costs them. You cannot price safely without this.
2. Find the value metric. Pick the unit the customer would happily pay more for as they get more value: a resolved ticket, a generated asset, a qualified lead. Price should rise with the value they receive, not just the compute you burn.
3. Set a floor that protects margin. Whatever model you choose, make sure your heaviest plausible user still leaves you with a defensible margin. Build in caps, fair-use limits, or overage tiers.
4. Make it legible to the buyer. The model has to be defensible inside the customer's own organization. A SMB owner wants one number per month. An enterprise wants a line item it can budget. If the buyer cannot explain your pricing to their boss, it does not matter how elegant your unit economics are.
5. Watch TCO, not the trial. The most expensive buyer mistake is choosing on the trial-month price instead of the 36-month total cost of ownership. The reverse is true for you as the seller: the cheapest headline price often hides the worst long-run margin.

Conclusion

The old SaaS instinct was to price for growth and trust that margin would take care of itself, because serving one more user cost nothing. That instinct is now dangerous. With AI, serving one more query costs something real, and it scales with success. The teams that win in 2026 are not the ones with the cleverest pricing page. They are the ones who know their fully loaded cost per transaction cold, drive it down with caching and routing, and pick a model that lets the customer's bill grow in step with the value they get. Price the cost you can actually see, govern the usage you can't, and never let your best customer be your worst margin.