Hybrid pricing combines a predictable base fee with a usage component. It has become the default structure for AI products because it gives buyers a budgetable number and gives the vendor a way to recover cost from the heaviest users. The hard part is choosing the three numbers. This guide derives them from cost data rather than competitor pricing pages.
Step 1: measure your blended cost per unit
Start from real traffic. Suppose your assistant runs on GPT-4.1 at $2.00 per 1M input tokens and $8.00 per 1M output tokens, and a typical call carries 9,000 input tokens and 700 output tokens. One call costs $0.0236. Divided by the 9,700 tokens it uses, the blended cost is $2.43 per 1M tokens.
Measure this per feature, not per product. A summarization feature with long inputs and short outputs has a very different blended cost from a drafting feature. If your features differ by more than a factor of two, consider metering in a unit that hides the difference, such as credits per action, while you account in tokens internally.
Step 2: choose the unit the customer sees
Tokens are precise and unfamiliar. Many buyers prefer a unit tied to their work: documents processed, conversations resolved, agent runs or minutes. The rule of thumb we use is to expose tokens to technical buyers and work units to business buyers, and to keep a fixed, published conversion between the two so finance can audit the bill.
Step 3: size the allowance
The allowance should cover the median tenant comfortably and the 75th percentile tenant most months. Pull a month of per-tenant usage and read off those percentiles. In this example the 75th percentile tenant uses about 20M tokens, so the allowance is 20M tokens, which costs 20 x $2.433 = $48.66 at list price.
An allowance that most customers exceed is just a lower base price with a surprise attached. An allowance nobody comes close to is a missed price increase.
Step 4: set the base fee from the margin target
If your gross margin target is 75% and the allowance costs $48.66 at most, plus other per-tenant delivery costs, the base fee must be at least four times that total, because a 75% margin leaves 25% of revenue for cost. With $199 as the base fee, a tenant who uses the full allowance has a model-cost margin of ($199 - $48.66) / $199 = 75.5%, which clears the target before hosting and support are allocated.
Step 5: price the overage
Overage should never be cheaper per unit than the included usage. To hold the same 75% margin it needs a 4x markup on blended cost: 4 x $2.433, rounded up to $9.80 per 1M tokens, a margin on overage of 75.2%. Round to a number buyers can remember and publish it.
If that looks high next to provider list prices, remember that overage carries no extra hosting or support allocation, so many teams accept a lower margin on overage than on the base plan, typically 50% to 60%. What matters is that the blended margin across base and overage stays on target.
If you route some traffic through cache or fall back to cheaper models, your real blended cost is lower than list. Keep the overage price fixed anyway; the saving is your margin, and changing customer prices every time your cost moves erodes trust.
| Component | Price | Your cost | Margin |
|---|---|---|---|
| Base fee incl. allowance, fully used | $199 | $48.66 | 75.5% |
| Overage per 1M tokens | $9.80 | $2.43 | 75.2% |
Step 6: protect both sides
Two guardrails make hybrid pricing safe to sell. For the customer, an optional monthly spend ceiling, so an agent loop on their side cannot produce a bill they never intended. For you, a circuit breaker in the request path, so the same loop cannot produce a provider invoice before the customer's ceiling even registers. Both depend on metering that is current to the request, not to the end of the day.
In PALM SOUQ, the allowance and overage live on the tier policy, the spend ceiling is a hard cap on the tenant, and usage events flow to Stripe, Metronome, Orb or Lago with the same tenant ID your billing system already uses.
Review cadence
- Monthly: share of tenants exceeding the allowance, and overage revenue as a share of total.
- Quarterly: blended cost per unit after model changes and cache improvements.
- Twice a year: allowance size and overage price against the current usage distribution.
Vineet Saresar
Founder & CEO, PALM SOUQ
Founder of PALM SOUQ. Sets company direction, product strategy and customer partnerships, and leads the work to make runaway LLM spend impossible for AI SaaS teams.