For a $30 seat to be worse value than paying per token, a person has to send about 166 messages a day. Not 35, which is what a busy knowledge worker actually sends. A hundred and sixty-six — one every two and a half minutes across an entire working day, every working day. That single number is what this calculator produces, and once you have seen it the seats-versus-API argument usually stops being an argument.
Here is where it comes from. Twenty-five people, $30 a seat, is $750 a month. At 35 messages each across 21 working days that is 25 × 35 × 21 = 18,375 messages. Each carries 1,800 input tokens — the question plus the conversation history resent every turn — and returns 500 output tokens. On Claude Sonnet 5 at $2.00 input and $10.00 output per million:
1,800 ÷ 1,000,000 × $2.00 = $0.0036, plus 500 ÷ 1,000,000 × $10.00 = $0.0050, giving $0.0086 a message.
18,375 × $0.0086 = $158.03 a month of tokens against $750 of seats. Divide $750 by $0.0086 and you get 87,209 messages, which spread over 25 people and 21 days is 166 messages each per day. Below that the API is cheaper on tokens. Above it, seats are.
Why the break-even is expressed per person per day
Because that is the only form of the number anyone can sanity-check. Told that the break-even is 87,209 messages a month, nobody has any intuition about whether that is a lot. Told that every person on the team would need to send 166 messages every working day, everyone knows immediately — that is not what happens.
It also makes the comparison portable. Seat prices and team sizes change, and the monthly totals move with them, but "how many messages a day does each person send" is a fact about your organisation you measure once and reuse. Pull it from your logs; the gap between what people say they send and what they send is consistently large in both directions.
Two things move the break-even meaningfully. Prompt caching is the first: Claude Sonnet 5 reads cached input at $0.20 per million against $2.00, and since most of those 1,800 input tokens are a resent conversation and a fixed system prompt, caching cuts the message to $0.0054 and pushes the break-even out to 266 messages per person per day. The second is the model tier. Run the same traffic on Claude Haiku 4.5 at $1.00 and $5.00 and a message costs $0.0043, doubling the break-even again. Cheaper models do not make the API win; they make the seat licence look even worse per token, which is a different claim.
The build cost is where the comparison gets honest
The calculator's default build-and-run figure is $900 a month, and it changes the answer completely: $158.03 of tokens plus $900 of running your own client is $1,058.03, against $750 of seats. At 25 people the API loses despite costing a fifth as much in tokens.
Set that field to zero and you are asserting that a chat interface, authentication, conversation storage, file upload and parsing, web search, mobile and desktop clients, admin controls, usage reporting, retention settings and a support rota all cost nothing to build and nothing to keep running. For a team that already operates an internal client and is adding one more use case to it, zero is fair. For a team that does not, it is the single most optimistic number on the page.
That is what a seat licence actually buys. The tokens are the cheap part — $158 of the $750 in this example. The other $592 is software you did not write, an on-call rota you do not staff, and a vendor whose job is to keep it working. Whether that is good value depends on whether you would otherwise have built it, not on the token arithmetic.
Scale changes the verdict because the build cost is fixed while the seat bill is not. Take the same usage at 200 people: seats are $6,000 a month, while metered usage is 200 × 35 × 21 × $0.0086 = $1,264.20, plus the same $900, for $2,164.20. The API is now $3,836 a month cheaper. Somewhere between 25 and 200 seats the lines cross, and the calculator will tell you exactly where for your own numbers.
Who subsidises whom
Under seat pricing, light users subsidise heavy ones. The person who sends five messages a day consumes 5 × 21 × $0.0086 = $0.90 of tokens and pays $30. The person who sends 200 consumes $36.12 and also pays $30. Roughly speaking, forty quiet users fund one power user, and the vendor prices the seat expecting exactly that distribution — which is why per-seat pricing is generous to the enthusiasts and terrible value for everyone who opens the tool twice a week.
Metered pricing inverts it perfectly. Every user costs what they use; nobody subsidises anybody. This sounds fairer and it introduces a problem seats do not have, which is that your bill now moves with behaviour you do not control. A team that discovers a genuinely useful workflow can triple your spend in a fortnight, and the finance conversation that follows is about a variable line item rather than a headcount multiplication. Plenty of organisations pay a real premium for a bill they can predict, and that preference is legitimate rather than innumerate.
The hybrid is usually the right answer and rarely the one on the table: seats for the ten people who live in the tool, metered access behind your own interface for the eighty who use it occasionally, sized with the AI chatbot cost calculator.
What the arithmetic leaves out
Four factors sit outside this calculation and can outweigh all of it.
Data handling. Enterprise seat agreements typically come with contractual commitments on training exclusion, retention windows and regional processing. The same commitments exist on the API side but you have to configure them and prove you did. If you operate under GDPR, HIPAA or a client contract with audit rights, the paperwork attached to the seat is part of what you are buying.
Support. When the tool breaks at 09:00 on a Monday, a seat licence gives you someone to call. Your own client gives you a Slack channel of people asking why it is broken.
Access control and offboarding. Removing a leaver from a seat list is one click that is audited. Removing them from something you built is only as reliable as the auth you wired up.
Shadow spending. If you decline both options, people expense personal subscriptions and paste company data into consumer accounts. That is not a saving.
Measure your actual messages per person per day, set the build cost to something you would defend in a meeting, and read the break-even figure rather than the headline. If it lands above 100 messages a day, seats are the answer at your size, and the way to reduce that bill is fewer seats rather than a migration. Cut your LLM bill covers the levers that work once you are metered, and the LLM API cost calculator prices the traffic model by model.
Prices used by this calculator were last verified on 15 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.