Most AI SaaS companies launch with a seat price because that is what their buyers already understand. A seat price works when the marginal cost of a user is close to zero. LLM features break that assumption: the cost of a user is now a function of how many tokens their workflows consume, and that distribution is wide.
This post walks through the three pricing structures we see most often in PALM SOUQ accounts, with the arithmetic shown so you can replace our numbers with yours. All prices come from the list prices in the PALM SOUQ LLM Price Index.
Two tenants on the same $49 seat
Take two customers of a document assistant built on Claude Sonnet 4.5, which lists at $3.00 per 1M input tokens and $15.00 per 1M output tokens. The light customer sends 2M input tokens and receives 200k output tokens in a month. That costs 2 x $3.00 + 0.2 x $15.00 = $9.00. On a $49 seat the gross margin on that customer is 81.6%.
The heavy customer runs long agent sessions over large contracts: 60M input tokens and 4M output tokens. That is 60 x $3.00 + 4 x $15.00 = $240.00. Same seat, same invoice line, and the account loses $191.00 every month. One heavy account erases the margin of roughly 5 light accounts.
Option 1: keep the seat, cap the cost
The least disruptive change is to keep the seat price and enforce a cost ceiling per tenant. The customer sees no new line items. When a tenant reaches its ceiling, the gateway either degrades to a cheaper model or returns a clear limit message that the product can turn into an upgrade prompt.
This protects margin but it leaves revenue on the table. The heavy customer above is getting real value from those 60M tokens. Capping them at the seat level tells your best user to use the product less.
Option 2: credits
Credit systems sell a prepaid balance and draw it down per action. They make cost visible to the customer and they move cash collection forward. The difficulty is choosing the exchange rate. If one credit maps to a fixed action such as one document summary, your margin still varies with document length. If one credit maps to tokens, the customer has to reason about tokens, which most buyers do not want to do.
Credits work best when the action is uniform in size, or when you meter in tokens internally and translate to a unit the customer already understands, such as pages processed or agent minutes.
Option 3: allowance plus overage
The structure we recommend most often keeps the seat and includes a token allowance worth a fixed amount of COGS. Usage beyond the allowance is billed as an overage at cost times a markup. With a $15 allowance and a 1.3x markup, the light customer never sees an overage. The heavy customer pays ($240.00 - $15) x 1.3 = $292.50 on top of the seat, for a total of $341.50 against $240.00 of COGS.
That heavy account moves from a loss of $191.00 to a gross margin of 29.7%. The overage is predictable for the customer because it scales with the work they chose to do.
| Structure | Light tenant margin | Heavy tenant margin | Customer experience |
|---|---|---|---|
| Seat only | $40.00 | -$191.00 | Simple, unlimited |
| Seat with cost cap | $40.00 | Capped at seat | Limits or degraded model |
| Seat + $15 allowance + 1.3x overage | $40.00 | $101.50 | Pays for heavy use |
What you need before you can bill an overage
Every one of these structures depends on knowing cost per tenant in close to real time. An allowance is meaningless if you learn about the overage when the provider invoice lands. You need input, output, cache read and cache write tokens attributed to your customer ID on every request, priced at the model that actually served it, and pushed to your billing system as usage events.
That is the job PALM SOUQ does in the request path. The gateway reads the X-PalmSouq-Tenant header, prices the response from the provider usage block, and forwards the event to Stripe, Metronome, Orb or Lago. Quotas and fallbacks read the same ledger, so the cap, the allowance and the invoice always agree.
A practical sequence
- Meter first. Run two weeks of traffic through the gateway with no limits so you can see the real cost distribution per tenant.
- Set a soft cap at the 95th percentile of tenant cost and alert, without blocking anyone.
- Pick an allowance that covers your median tenant with room to spare, then price overage at cost plus the markup your gross margin target requires.
- Turn on billing sync and show customers their usage inside the product before the first overage invoice goes out.
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.