A $20 seat and a metered API key buy the same tokens in very different ways. The seat is a flat fee with limits; the API is a price per token with no limits and no floor. Which is cheaper depends on one number most teams never compute: calls per user per day at which the two cross.
The two pricing shapes
A subscription bundles usage into a monthly fee with rate limits, fair-use rules and, increasingly, a credit system for the most expensive models. The marginal cost of one more question is zero until you hit a limit. An API charges per token from the first token, so the marginal cost is constant and the total is proportional to usage. For a team of five people asking a few questions a day, the subscription usually wins by a wide margin. For a product that answers ten thousand customer messages a day, the subscription is not even an option — there is no seat for a machine.
The break-even formula
Let p be the monthly seat price, c the cost of one API call at your token mix, and
d the number of billed days. Then a seat costs the same as metered usage at
break-even calls per user per day = p / (c × d)
With a $20 seat, 30 days, and a call that costs $0.006 on Claude Sonnet 5.5, the break-even is 111 calls per user per day. Below that, the seat is cheaper; above it, the API is. The same formula rearranged gives the break-even seat price: if your API spend per user is $8 a month, a seat would have to cost less than $8 to beat it.
A worked team example
Five people, each sending 60 requests a day of 1,500 input and 400 output tokens with 30% cache hits, on a model at $2/$10 per million with a $0.10 cached rate:
- Per call: 1,050 × $2 + 450 × $0.10 + 400 × $10, divided by a million = $0.006145.
- Per user per month: $0.006145 × 60 × 30 = $11.06.
- Team of five: $55.31 against $100 of seats. The API is $44.70 cheaper per month.
Now raise usage to 120 calls a day per person and the API bill doubles to $110.60 while the seats stay at $100. The crossover sits at about 108 calls per user per day — close to the formula's prediction, and close enough that a team should re-run the numbers whenever its habits change.
What the subscription includes that the API does not
The comparison above is deliberately narrow: tokens for tokens. A seat also buys an interface, file handling, memory, connectors, a mobile app, and someone else's uptime. The API buys none of that — you build the product around it. For an individual, the seat is almost always the better deal once you count the engineering time to reproduce the app. For a company shipping its own product, the API is the only path, and the interesting question becomes margin rather than subscription versus API.
Margin math for a product built on tokens
If you resell model output, your gross margin per user is
margin = (price per user − token cost per user) / price per user
At $10 per user and $6 of tokens, the margin is 40% before infrastructure and support. Two levers move it: price, and tokens per user. Caching a shared prefix, trimming retrieved context, routing simple requests to a small model and batching background jobs are the four changes that most often take $6 down to $2. The calculator's subscription tab shows the API cost per user and the seat price that would match it, which is the same arithmetic read from the other side.
Where the model is fragile
Three assumptions deserve suspicion. First, average usage hides the tail: a team's median user may send 20 calls a day while its top user sends 400, and a seat-based price protects you from that tail while an API budget does not. Second, prices move: Gemini 4 Carbon doubles on 1 December 2026, Gemini 3.8 Flash doubles on 1 January 2027, and frontier prices have fallen roughly every generation — a break-even computed today is not a break-even for next quarter. Third, token cost is not the whole cost: a cheaper model that fails 15% more often can cost more per resolved ticket than a model with twice the price.
A short checklist
- Measure a real week of calls, input tokens, output tokens and cache hits per user.
- Compute cost per call on two models: the one you would choose for quality and the cheapest one that passes your evaluation.
- Find the break-even calls per day against your actual seat price, not the list price you remember.
- Add the tail: price the 95th-percentile user as well as the median.
- Re-run the comparison when a provider changes a price or you change a prompt.
That is the whole discipline. The calculator exists so steps two and three take a minute instead of an afternoon.