AI products broke the cleanest thing about SaaS economics. Classic SaaS had near-zero marginal cost per user, so flat subscriptions were nearly pure margin and pricing was a positioning exercise. An AI SaaS pays real money per action (every generation, every analysis hits a metered API), so your costs are usage-shaped whether or not your pricing is. Align them and you create one set of problems; don't and you create another.
This article maps the actual trade-off, including the failure mode that doesn't show up until you scale: the usage-pricing cliff.
The case for usage-based, and the cliff
Usage-based pricing has obvious appeal for AI products: revenue tracks cost (no whale on a $29 plan burning $400 of inference), entry friction is low, and growth is built in, since customers who get more value pay more without a sales conversation.
Then the cliff. The customer who gets the most value ends up the most price-sensitive, because their bill grows linearly with their success. Your best account watches an invoice climb month over month, and at some threshold the line item gets noticed by someone whose job is noticing line items. Then: pressure to negotiate custom rates, active usage-reduction efforts, or churn to a flat-priced competitor, precisely the accounts you least want shopping around. Usage pricing also fights you operationally: customers can't budget unpredictable bills (enterprise procurement actively punishes this), anxiety suppresses the exploratory usage that drives adoption, and your MRR becomes a forecast instead of a number.
The cliff is why the pure-metered model that looked inevitable for AI products in 2023 quietly lost to hybrids by 2026.
The case for subscription, and its own failure mode
Flat tiers give customers budgetable bills and you predictable MRR, preserve the carefree usage that makes products sticky, and keep the pricing page readable. For AI products, the failure mode is the margin invert: any flat price is a bet on average usage, and AI usage distributions are violently long-tailed. A handful of power users can take a tier from 85% margin to negative, and unlike hosting overages, inference overages scale fast. (Quotas and model tiering are the defenses, but they cap the damage rather than remove the mismatch.)
Flat pricing also leaves money on the table at the top: your heaviest users are getting outsized value at a fixed price, which is generous right up until it's unsustainable.
What actually works: the 2026 hybrid playbook
The models that survived contact with real customers all combine a predictable base with usage protection:
Tiers with included usage (the default answer). Subscription tiers, each including a generous usage allowance, with visible meters and a humane overage path: soft warnings, then a tier-upgrade prompt, only then metered overage. Customers budget; you're protected from the tail; the upgrade path is usage-driven and feels fair. This is the right starting model for most AI SaaS, and most of the products you admire run it.
Credits, for spiky workloads. Sell credit packs or monthly credit allowances when usage is naturally bursty (generation tools, batch processing). Credits convert anxiety into a purchasing decision, prepay your inference costs, and let heavy months coexist with quiet ones. Cost: some buyers hate the mental accounting; unused-credit policies need care.
Seat + usage, for team products. Per-seat base (predictable, maps to procurement) plus pooled team usage allowance. The B2B-friendly hybrid.
Two design rules whatever you pick. First, price the outcome unit, not the API unit: charge per document processed, per report, per campaign, never per token. Customers should buy your value in your vocabulary; tokens are your cost structure, not their benefit (pricing on value is still the law). Second, engineer your floor: model routing, caching, and quotas don't just cut costs; they're what makes generous included allowances affordable, which is what makes the whole hybrid feel abundant instead of stingy. Your margin structure is a pricing feature.
Choosing for your product
| Your situation | Start with |
|---|---|
| Typical AI SaaS, individual/prosumer | Tiers with included usage |
| Bursty generation workloads | Credits |
| B2B team product | Seat + pooled usage |
| Usage variance genuinely low | Flat tiers, quotas as backstop |
| Infrastructure/API product | Pure usage (your buyers budget this way) |
And the meta-rule: you'll reprice within a year, so optimize for learnability: visible meters teach you your real usage distribution, which is the data your next pricing iteration needs. Pricing isn't a launch decision; it's an ongoing experiment with revenue attached.
Frequently Asked Questions
What is the usage-based pricing cliff?
The cliff is the point where a usage-priced customer's growing bill triggers active resistance: because their costs scale linearly with their success, your most successful customers become your most price-sensitive ones, negotiating discounts, suppressing usage, or churning to flat-priced competitors. It's the structural flaw that moved most AI SaaS from pure metering to hybrid models with included allowances.
Should my AI SaaS use usage-based pricing?
Pure usage-based: probably not, unless you're selling infrastructure or APIs to buyers who already budget that way. For most AI products the 2026 default is subscription tiers with generous included usage and a humane overage path: it gives customers budgetable bills, protects you from long-tail inference costs, and creates a fair, usage-driven upgrade trigger. Reserve credits for genuinely bursty workloads.
How do I price AI features without losing money on heavy users?
Three layers: include a usage allowance in each tier sized so 90–95% of users never hit it, engineer your cost floor down (route routine calls to cheaper models, cache aggressively, set hard per-user quotas), and make the path past the allowance an upgrade prompt rather than a surprise bill. The combination keeps margins intact across the usage distribution while letting normal users feel the product is abundant.
Should I charge per token like the AI providers charge me?
No: tokens are your cost unit, not your customer's value unit. Price the outcome your product delivers (per document, per report, per generation, per seat) and translate internally. Customers can't predict or even understand token consumption, and exposing it imports your cost anxiety into their buying decision; the products that win price in the vocabulary of the job being done.