Plan token quotas that protect your margin
If your product passes LLM calls through to users, an unbounded user is an unbounded bill. Token-based quotas turn that risk into a fixed, plannable number: each tier gets a monthly token allowance, and your maximum exposure is simply the sum of those allowances times your token price. This planner computes the worst-case cost per tier and your total cost ceiling so you can price every plan above what it can ever cost.
How it works
For each tier, the maximum monthly cost per user is:
max_cost_per_user = tier_tokens / 1,000,000 × blended_price
tier_ceiling = max_cost_per_user × users_in_tierSummed across tiers, the total cost ceiling is the most you could spend on inference in a worst-case month. The planner also reports a blended cost per user across your whole base, which is the number to keep below your blended revenue per user. As long as each tier’s price exceeds its per-user ceiling, the unit economics hold no matter how heavily users engage.
Worked example
Imagine a SaaS product with three tiers and a blended token cost of $2.00 per million tokens:
| Tier | Monthly tokens | Users | Max cost/user | Tier ceiling |
|---|---|---|---|---|
| Free | 50,000 | 1,000 | $0.10 | $100 |
| Pro at $19/mo | 2,000,000 | 200 | $4.00 | $800 |
| Enterprise at $99/mo | 10,000,000 | 20 | $20.00 | $400 |
| Total | 1,220 | $1,300 |
For example, the Pro tier charges $19/month and the maximum inference cost per Pro user is $4.00 — that is a $15 gross margin per Pro user at full quota usage. In practice, most Pro users will consume a fraction of their allowance, so actual margins are higher. But the plan guarantees positive unit economics even in the worst case.
The Free tier costs at most $100/month across 1,000 users — a manageable acquisition cost that is worth covering if free users convert to Pro at a reasonable rate.
Setting the blended token price
The blended price is your effective per-million-token cost across all models you serve, weighted by the mix of input and output tokens. For example:
- If your calls are 80% input and 20% output, and input costs $3/M while output costs $15/M:
- Blended = (0.8 × $3) + (0.2 × $15) = $2.40 + $3.00 = $5.40 per million tokens
Prompt caching can reduce the effective input cost significantly — if you cache system prompts heavily, the blended price can drop by 50–80% of the input component.
Tips for token SLAs
- Price above the ceiling, not the average. Averages feel safe until a power user or an abuse spike arrives; the ceiling is what an SLA must cover.
- Make quotas visible. Show users their remaining tokens so the limit is a feature, not a surprise wall.
- Offer overage, don’t just block. Metered overage above the quota recovers cost from heavy users instead of cutting them off.
- Re-blend the price when models change. Switching default models or enabling caching shifts your effective per-token cost — re-run the plan when it does.
- Review free tier cost as a share of CAC. The free tier inference cost is a customer acquisition cost. As long as it is well below the LTV of a converted user, it is a legitimate investment.