Key Takeaways

• Every AI pricing model resolves to a choice about correlation: how tightly the invoice tracks inference cost. Per-token correlates almost perfectly, flat subscription not at all.

• The models aren't mutually exclusive. Anthropic runs per-token rates, a per-hour runtime rate, a per-call search rate, and a free-allowance-plus-overage rate in the same price list.

• Real published rates span four models: $2 per million input tokens for Claude Sonnet 5, $0.08 per session-hour for Managed Agents runtime, $10 per 1,000 web searches, and 1,550 free container hours a month before $0.05 per hour.

• Hybrid is the working default for shipped AI products: a platform fee, metered consumption on top, and a credit balance that absorbs variance.

• Changing model is a migration, not a config change. Existing contracts, mid-cycle proration, and grandfathered plans all have to keep working while the new model runs alongside.

What are the main types of AI pricing models?

Six structures cover almost everything shipping today, and each one puts a different unit on the invoice.

Per-token and per-request are the closest to raw cost. Credit-based sits on top of any of them as a payment mechanism rather than a competing unit, which is why teams often describe their pricing as credit-based and per-token at the same time without contradiction.

Which AI pricing model should you pick?

Pick the model that matches how your cost behaves, then adjust for what your buyer can forecast.

The decision splits cleanly along two questions:

• Does your cost scale with use? If yes, a metered model protects margin. If no, a subscription captures more value per customer.

• Can the buyer predict their own volume? If no, pure metering makes the purchase hard to approve, and a credit balance or a committed floor fixes it.

Enterprise buyers usually need a committed number for procurement, which pushes toward a subscription floor with metered overage. Self-serve buyers tolerate pure metering because there's no approval cycle to survive.

What breaks when you change an AI pricing model?

Plan migration breaks first, because existing customers keep their old terms while new customers land on the new ones.

The pieces that need to keep working through a change:

• Proration for anyone who switches mid-cycle, so nobody pays twice for the same period

• Grandfathering, since legacy accounts often stay on old rates indefinitely by contract

• Historical invoice accuracy, because reissuing a corrected invoice under a rate that didn't exist yet fails an audit

• Metering continuity, as a new model frequently needs a unit the old one never measured

That last one is the trap. Moving from per-seat to per-token means you need token counts for periods you never metered, and you can't backfill data you didn't collect.

How do teams run AI pricing models in production?

Teams that change pricing often separate the pricing configuration from the application code, so a new model ships without a deployment.

Flexprice is enterprise-grade, open source usage based billing infrastructure for AI and SaaS companies. It can be deployed in your own VPC, on-prem, or on Flexprice's managed cloud. Simplismart runs 750+ pricing features on it and iterates on pricing 6x faster than before, with Shubhendu Shishir, their Head of Engineering, describing pricing as a continuous growth lever rather than a quarterly project. The configuration for each model lives in Pricing Models rather than in a service that needs redeploying.

If you're weighing a change to your own model, book a demo and walk through the migration path for your existing contracts.

Related reading

Further reading on choosing and changing a pricing model:

• saas pricing models covers the SaaS structures these AI models evolved from.

• hybrid pricing guide goes deep on combining a fee with metered usage.

• what happens to billing when you change your pricing model is the migration problem in detail.

• how to test usage based pricing before fully commiting covers piloting a model before a full switch.

• modern pricing infrastructure for ai companies 2025 describes the stack these models run on.

FAQ

Is per-token pricing better than subscription pricing for AI?

Per-token pricing protects margin better; subscription pricing sells better. Tokens track your cost almost exactly, so no customer can be unprofitable. A subscription gives the buyer a predictable number and gives you more revenue from light users. Most teams end up with both, using a subscription floor for predictability and token overage for protection.

Can you combine credits and usage in one AI pricing model?

Yes, and it's one of the more common shipped combinations. Usage metering measures what was consumed, and the credit balance is what the consumption draws down. The credits become the payment mechanism and the spending cap at once, while the usage model still decides the rate at which they deplete.

How is AI pricing different from SaaS pricing?

Cost of goods varies per customer in AI and barely varies in SaaS. That single difference cascades: AI pricing needs metering, per-customer margin visibility, and a stop mechanism, none of which seat-based SaaS pricing requires. It also means engagement can hurt an AI product's margin, which reverses the usual read on usage growth.