AI SaaS pricing models: what actually works under $2M ARR
There are five pricing models for AI products: per-seat, flat subscription, usage-based, credit-based, and outcome-based. Most AI SaaS in 2026 converge on a hybrid: a base fee that covers typical usage, plus credits or overage for heavy use. The right one for you depends on one number you can pull from your own logs today: how much more your 90th-percentile customer costs to serve than your median customer.
Almost everything written about AI pricing is written by billing vendors and VCs for companies with a pricing team. This guide is for the founder deciding alone, with a Stripe account, an OpenAI or Anthropic invoice, and no one to delegate the decision to. Every model below comes with the data check you should run before picking it.
Why did AI break SaaS pricing?
Traditional SaaS has near-zero marginal cost, so any pricing model that grows with adoption works. AI SaaS has a meter running on every interaction: tokens, GPU seconds, vector queries. Your cost per customer is no longer roughly constant, it is a distribution with a long tail. A flat fee against a long-tailed cost curve is a bet that your heavy users will stay light. In 2020 a16z documented AI companies running 50 to 60 percent gross margins against the 60 to 80 percent SaaS norm, and the agent era has made the spread worse, not better.
The public failures all share this shape. GitHub Copilot reportedly lost an average of $20 per user per month in its flat-fee era, with some users costing $80 against a $10 price (ARK Invest disputed the figures). Replit's gross margin swung from 36 percent to negative 14 percent in roughly two months after its agent launched. Cursor repriced in June 2025 and the rollout went badly enough that the CEO publicly apologized and issued refunds. None of these companies had a cost problem first. They had a pricing structure that did not meter the thing that cost them money.
What are the five AI pricing models?
| Model | How it works | Works when | Breaks when |
|---|---|---|---|
| Per-seat | Price per user per month | Usage per seat is uniform and value tracks headcount | Agents decouple work from headcount; one seat burns 50x another |
| Flat subscription | One price, unlimited use | Cost per customer is small relative to price | The P99 customer arrives |
| Usage-based | Metered per token, call, or minute | Cost and value scale together, buyers accept variable bills | Bill shock erodes trust; revenue becomes hard to forecast |
| Credit-based | Prepaid credit pool, each action burns credits | You need metering without exposing token math | Credits are opaque; customers feel gamed when burn rates change |
| Outcome-based | Price per result (per resolution, per qualified lead) | The outcome is crisply definable and attributable | Disputes over what counts; only works for narrow job-shaped products |
Per-seat and flat: the incumbents
Before you keep per-seat pricing, check whether cost per seat is actually uniform. Pull 90 days of usage, group by account, and compare the 90th-percentile cost per seat to the median. Under 3x, seats are fine and simple wins. Over 5x, your heaviest accounts are being subsidized by your lightest ones, and the subsidy grows as your best users engage more. That is the exact trap the Copilot numbers describe.
Usage-based: honest but scary
Pure usage pricing keeps your margin flat by construction, and buyers increasingly expect it for API-shaped products. Its failure mode is psychological, not financial: unpredictable bills make customers preemptively throttle their own usage, which caps your revenue at exactly the moment the product is most valuable. If you go this way, ship spend caps and alerts on day one. Replit's move to effort-based checkpoint pricing, where the price is visible only after the work is done, shows how much goodwill you burn when the meter is opaque.
Credits: the compromise everyone is converging on
Credits are usage-based pricing wearing a friendlier interface. The customer reasons in credits, you reason in tokens, and the exchange rate is your margin lever. Adoption is moving fast: credit-based pricing grew 126 percent year over year in 2025 among the top 500 companies in the PricingSaaS index. The failure mode is trust. If you re-rate credit burn without warning, you inherit the Cursor backlash. Publish the burn table and version it.
Outcome-based: powerful, narrow
Intercom charges $0.99 per resolution for Fin, Zendesk charges $1.50 per automated resolution, HubSpot dropped to $0.50 per resolved conversation in April 2026. Outcome pricing is the cleanest value story in AI, and the hardest to operate: you eat every dispute about what counts as a resolution, and Fin's own definition (which includes customers who simply go quiet) shows how contested the unit gets. Consider it only if your product completes a whole, countable job.
Hybrid: the default answer
A base subscription sized to cover your median customer, plus included credits, plus overage or top-ups past the cap. The base fee keeps revenue predictable, the meter keeps the P99 customer from eating your margin, and the included allowance keeps light users from feeling nickeled. Stripe's guide to AI pricing reports that 56 percent of AI company leaders use hybrid pricing, and Cursor, v0, Claude, and ChatGPT Business all ship some variant of it. This is also why our own seats versus usage versus credits comparison ends with two hybrid structures rather than a single winner.
How do you choose? Run the cost-spread test
You do not need a pricing consultant to get the shape right. You need three numbers from your own last 90 days: the median (P50), 90th percentile (P90), and 99th percentile (P99) of AI cost per customer.
- P90 / P50 under 3: your usage is uniform. Seats or flat pricing are fine. Revisit quarterly.
- P90 / P50 between 3 and 10: hybrid. Set the included allowance near your P75, price overage at 3x to 5x your marginal cost.
- P90 / P50 over 10, or P99 customers costing more than they pay: you have margin bombs today. Cap or meter heavy usage before you do anything else, then redesign around a hybrid or usage model.
If you have never computed cost per customer, that is the first fix. Our guide on calculating LLM cost and margin per customer walks through it step by step with a worked example.
What are AI companies actually charging in 2026?
The market has already voted. GitHub Copilot moved every plan to usage-based billing on June 1, 2026. Cursor sells subscriptions whose price equals an included usage pool, with token-rate overages. ChatGPT Business keeps seats but sells credit top-ups on top. The direction is one-way: flat and pure-seat models are gaining meters, and nobody with real inference costs is moving toward unlimited. The full breakdown, with sources and current prices, is in how 8 real AI products price in 2026.
Frequently asked questions
What pricing model should my AI startup use?
Default to hybrid: base fee covering typical usage, credits or overage past a cap. Only deviate when your data says usage is uniform (seats) or your product completes a countable job (outcome-based). Run the cost-spread test above before deciding.
Is per-seat pricing dead for AI products?
No, but it is being retrofitted with meters. Seats survive where usage per seat is predictable. For agentic products, where one user can consume 10x to 50x the compute of another, pure seats quietly become a subsidy for your heaviest accounts.
How much should I mark up LLM API costs?
Common guidance is 3x to 5x model cost, which puts inference at roughly 20 to 35 percent of the revenue it supports. If a feature's model bill exceeds a third of the revenue it generates, reprice the feature before optimizing the prompt.
Should my AI product have a free tier?
Only with a hard usage cap, because free users carry real marginal cost. Size the free allowance so that its total monthly cost is a marketing line item you would happily pay, then measure conversion against it.
Stop guessing what to charge
AI Pricing Lab rebuilds your pricing from your real Stripe and usage data: a per-customer margin map, three pricing models simulated on your last 6 to 12 months, and the migration pack to ship the change. Delivered in 10 business days. One-time fee, no subscription.
See the AI Pricing Reset$1,500 founding rate · first 10 AI SaaS · $2,995 list