Step one: measure cost per action, then set the mapping
- Tag every model call with its feature, per tracking AI costs per user, and let production tell you the real cost distribution per action: the summary averages $0.003, the document analysis $0.03, the agent workflow $0.20 with a wild tail.
- Define the credit’s cost basis, for example 1 credit is about $0.01 of model spend, and map actions proportionally: summary 1 credit, analysis 3, workflow 20. Round for friendliness, but keep the ratios within reach of reality, a flat 1 credit for everything invites users to arbitrage your expensive actions.
- Sell the credit at a multiple of the basis, commonly 3x to 10x raw model cost, sized so the p95 user’s mix still clears margin after your other costs, the same tail first pricing as usage based billing for AI products.
The drift problem: your exchange rate rots
Everything under the credit moves while its price stands still. Prompts grow, output drifts longer, models get swapped, and providers reprice, an 80 percent cut in July 2026, increases elsewhere per LLM price history. Teams discover the drift at the margin review. The fix is continuous: real cost per credit consumed, per feature, from live traffic, with the action mapping adjusted when the basis moves. Adjust credits per action, not the credit’s headline price; users forgive “analysis now costs 4 credits” far more readily than a currency devaluation.
// the health metric to watch, per feature:
real_cost_per_credit = measured_model_cost / credits_charged
// basis was $0.01. If analysis drifts to $0.018 per credit, remap the action.Rollover, refunds, and the ledger’s blind spots
- Rollover is deferred cost exposure. Banked credits are future model spend at future prices. If you allow it, cap the bank.
- Failed calls and retries: decide explicitly what spends a credit. Users forgive a lot except paying for your errors.
- The ledger lies sometimes. Bugs, abandoned streams, and mispriced actions spend tokens without spending credits correctly. Back the whole system with a per user dollar cap enforced before the call, per per user spending caps: whatever the credits say, no account can cost more than its ceiling.
Weckr is the measurement half of this article as a service: real cost per user and per feature at current rates, margin against what each user pays, and the dollar caps behind the ledger, all from two lines around your existing client. Credit systems built on top of it start from measured basis instead of guesses.
FAQ
How should credits map to AI actions in a SaaS?
Price actions in credits proportionally to their real token cost, then round to friendly numbers: if a short summary costs you $0.003 and a long document analysis costs $0.03, the analysis should cost roughly ten times the credits. Flat one credit per action across wildly different costs is the classic mistake: users optimize toward your expensive actions and your effective margin varies 10x by usage mix.
How do I price a credit?
Set the credit’s underlying cost from measured data (say 1 credit is roughly $0.01 of model spend at current rates), then sell credits at a multiple that covers your other costs and the heavy tail, commonly 3x to 10x the raw model cost depending on your market. The key discipline is knowing the real cost per action from production measurements, not estimates, because that number drifts.
Should unused credits roll over?
Product decision with a cost angle. No rollover keeps liability clean and matches the monthly cost pattern of your provider bills. Rollover is friendlier and creates deferred cost exposure: banked credits are future model spend at future prices, which cuts both ways as prices move. If you do roll over, cap the bank, and remember a price cut like July 2026’s makes banked credits cheaper to honor, while output drift makes them dearer.
What happens when a model price change breaks my credit economics?
Your credit price stands still while its cost basis moves, so margins drift silently, in your favor on cuts, against you on prompt growth and output drift. The defense is instrumentation: track real cost per credit consumed, per feature, continuously. When the basis moves more than some threshold, adjust the credits per action mapping rather than the headline credit price, users tolerate action repricing far better than currency devaluation.
Do I still need spending caps if I have credits?
Yes. Credits bound intended usage; caps bound reality. Bugs, retries, streaming abandonment, and mispriced actions all spend tokens without spending credits correctly, and a runaway loop can burn dollars fast while the credit ledger thinks everything is fine. A per user dollar cap behind the credit system is the invariant that catches every accounting mistake: whatever the ledger says, no user can cost more than X.
Keep reading
Set the rate from data. Check it from data.
A credit system is only as honest as its cost basis, and the basis is a measurement, not a constant. Weckr keeps it measured continuously, free for 50,000 requests a month. See per feature costs on the live demo, or start with the AI cost and margin guide.