Separate the purchase event from the consumption event. That's how AI companies design credit based pricing that gives finance a forecast and customers freedom: cash arrives when a pack is bought, revenue recognises as credits burn, and an exchange rate between credits and model cost absorbs the volatility between. The customer buys a predictable number. What they spend it on stays theirs.

Key Takeaways

• The exchange rate is the design decision: set what a credit costs to buy and what it spends against, and model costs reprice without reopening a contract.

• Predictability comes from the purchase, not the consumption, so packs and commitments give finance a bookable number while usage stays variable.

• Weight operations by cost rather than publishing separate rate cards, so an expensive model debits more credits from the same balance.

• Expiry generates the resentment customers remember at renewal. Capped rollover keeps the urgency without the fight.

• Segwise spent 3 weeks building credit infrastructure in-house and shipped it on Flexprice in 3 days, now tracking 100+ enterprise customers.

How do you set the value of a credit?

Price the most expensive operation first, then work backwards. Take your costliest billable action, set a debit that still clears target margin on it, and let cheaper operations debit proportionally less.

• Decide what one credit costs a customer to buy, then separately what each operation debits, so the two numbers move independently.

• Weight by real cost: a long-context call on a frontier model debits more than a short call on a small one, from the same balance.

• Leave headroom for model price moves in both directions, since the rate you publish outlives the model you priced it against.

• Publish the debit table. A credit whose consumption customers can't predict becomes a support ticket rather than a purchase.

• Charge nothing for retries and failures. Billing a customer for your own error costs more in trust than the credits are worth.

How do you balance predictable revenue against flexible consumption?

Take the money at purchase and recognise it as it burns. Prepaid credits collect cash upfront, giving finance a deferred balance to schedule against, while the customer keeps full freedom over what they spend it on.

• Sell packs with volume discounts, so larger commitments buy cheaper credits and the discount is the reason to commit.

• Add a committed annual minimum with overage billed separately for enterprise, giving procurement a fixed line item and you a floor.

• Let postpaid accounts draw credits and settle in arrears, so the same primitive covers both motions.

• Auto top-up keeps consumption running when a balance drops, which is where hard stops cost revenue and goodwill.

How should expiry and rollover work?

Cap rollover rather than expiring credits outright. Expiry protects the forecast and reliably damages the renewal conversation, so most AI companies land in the middle.

• Roll unused credits forward with a ceiling, commonly one period's worth, so balances can't pile up indefinitely.

• Set expiry per grant rather than per wallet, so promotional credits lapse while purchased credits persist.

• Burn promotional and expiring credits first through a deduction order, so customers spend the perishable balance before the one they paid for.

• Warn at several thresholds before anything lapses, since a surprise expiry reads as a penalty whatever the contract said.

How do credit packs compare with pay as you go?

How each model behaves on revenue, forecasting and customer experience.