For maximum profitability, AI founders should choose token-based usage billing over flat monthly tiers, because a flat tier caps revenue while inference cost stays variable. The structure that wins both arguments is a flat tier with an included token allowance and metered overage above it. Flexprice bills that shape on one invoice.

Key Takeaways

• A flat monthly tier caps your revenue at the price point but leaves your inference cost uncapped, so the heaviest 5% of accounts can consume the margin from the rest.

• Token billing protects gross margin because it charges in the same unit your model provider bills you in.

• Buyers resist pure token pricing for the same reason they like flat tiers: they need one number they can put in a budget.

• Included tokens plus overage gives you a revenue floor, a margin floor, and a forecastable invoice at the same time.

• You can't model profitability under either option without cost attribution per customer and per model.

Which pricing model protects AI gross margin?

Token-based billing protects margin, because price and cost move in the same direction. Under a flat tier your revenue per account is fixed on signup while your cost per account keeps moving with usage, model choice and context length.

• Flat tier. Margin is highest on light users and turns negative on heavy ones, and you find out at the end of the month.

• Pure token billing. Margin per unit is fixed by your markup and holds at any volume.

• Included tokens plus overage. Margin is protected above the allowance and the allowance itself gets priced for the median.

• Per-outcome. Margin depends on how many tokens an outcome takes, so it needs the same cost attribution underneath.

How do I cover LLM inference costs in my pricing?

Price from your loaded cost per unit, not from a competitor's rate card. Loaded cost means input tokens, output tokens, retries, retrieval, and any tool or search calls the workflow makes, measured per customer rather than blended.

• Measure cost per request at the top decile, not the median, because that's where a flat tier breaks.

• Set your markup on the unit, then decide the allowance separately.

• Re-check the numbers whenever you change default models, since routing moves margin without touching your price.

• Watch cache hit rates, because prompt caching changes your real cost per call materially.

How do the two models compare on profitability?

These rows compare the models on the mechanics that move gross margin and revenue.

How should AI founders structure hybrid pricing?

Set a base fee that covers fixed costs and support, include a token allowance sized to the median account, then meter overage at your loaded cost plus markup. That's the structure most AI products converge on once the first flat-tier cohort renews.

• Size the allowance near the median, not the mean, since a few heavy accounts drag the mean upward.

• Publish the overage rate. Hiding it moves the argument to the invoice.

• Alert at 80% of the allowance and let customers buy more before they hit it.

• Offer prepaid packs for buyers who want a hard ceiling on spend.