Ten people at USD 20 a month is USD 2,400 a year. In Vietnam that pays two months of a fresh graduate's salary. So before you pay, answer one question: of those ten people, how many actually opened an AI tool last week?
You are paying for 'available anytime', not for AI itself
A seat is billed per head: ten people, ten charges, whether they use it or not. Tokens are billed per usage: no call, no charge, and cost moves with workload. Neither model is wrong. The mistake is treating headcount as usage.
A seat buys 'available anytime'; a token buys 'how much was used'. The gap between them is what you paid for conversations that never happened.
Count heads first: the 30% rule
We have looked at a lot of team dashboards, and the pattern is stable: in a ten-person team, no more than three open an AI tool every day. The rest use it twice a week, or only during a deadline week, yet the seat fee is charged all twelve months. That is the 30% rule.
Counting takes no tooling: open the usage history in your billing dashboard once, or ask in the group chat 'who used AI last week'. You will have the answer in five minutes.
Run the numbers
The break-even formula is one line:
Monthly break-even usage = total seat cost per month ÷ your token unit price
Plug in ten ChatGPT Plus seats: USD 20 × 10 = USD 200 a month. Say you use Qwen 3.7 Max, listed at USD 1.65 per 1M input tokens and supplied by LOCOM at USD 0.96 per 1M:
- 200 ÷ 0.96 = 208M input tokens per month
- Divided by ten people = 20.8M each per month
- Divided by 22 working days = 947k tokens each per day
How much is 947k tokens? Roughly reading 700,000 words of documents every day. Nobody uses AI like that. Switch to DeepSeek V4 Flash at USD 0.09 per 1M and the break-even point becomes 2.2 billion tokens a month - even further out of reach.
The reverse also holds: if your team really does cross that line, seats are the cheaper deal. The formula does not take sides - it only tells you where the line sits.
Three kinds of teams, three answers
A - The whole team uses it daily and cannot give up GPT's high-end reasoning. Buy seats, but only for the genuinely heavy users. Let tokens cover the rest; a year of savings pays for half an intern.
B - Three heavy users, seven occasional ones - the most common shape. Go hybrid: three seats to keep the daily experience intact, tokens for everything else. This is the optimum for almost every ten-person team.
C - Project-based or seasonal work. Pure tokens. Peak season can triple your usage at no extra charge, and off-season drops to almost nothing - a seat fee cannot do that.
The hidden cost: multiple models
Marketing copy wants GPT, code wants Claude, Vietnamese content wants Qwen and DeepSeek - models trained on Chinese corpora that actually handle Vietnamese technical terms more accurately. Seats mean buying a full round of seats for each model, tripling the cost. Tokens mean one API, twenty-plus models behind a single key, billed by calls.
Few teams count this line item, yet it is the real dividing line for multi-model teams.
When you should just buy the seats
In fairness, tokens are not universal. If you want the ChatGPT app, voice mode, Canvas or web search - the interface features - seats really are simpler. Tokens are an API and must be wired into your own tool first. If nobody on the team writes code, just buy the seats; do not trade efficiency for savings.
The deciding question is simple: are you buying an interface you click open, or a capability you wire into a system?
Back to that USD 2,400. It is not an 'AI cost' - it is a test of how hard your team actually uses the tool. Count heads first, find the line second, open the wallet last. Reverse the order and the money leaves first.
FAQ
Can a token bill suddenly explode in one month?
No. Token billing is linear - double the usage, double the cost - and you can set a monthly cap in the dashboard that stops calls when reached. The real blowout is on seats: you pay for ten people, three use it, and those seven are the actual waste.
How far behind GPT are Qwen and DeepSeek?
To be clear: complex reasoning, long code chains and nuanced multilingual writing still favour GPT and Claude, at several times the price. But daily copywriting, translation, customer replies and data cleanup - the high-frequency work - are well served by Qwen and DeepSeek, which handle Vietnamese technical terms more steadily. The test is not 'which is stronger' but 'how strong your task needs to be'.
Can a non-coding team still use tokens?
Yes, but you need a front end. Use an existing third-party client that connects to the API, or have one person spend half a day setting up a shared entry point. If even half a day is not available, buy seats first; a tool only counts once it is actually in use.
What is the first step?
Count heads. Open the usage history once, or ask 'who used it last week', and write the number down. If it is not more than thirty percent of the team, route the saved seat budget through tokens for one month as a control - two invoices will put the answer on the table.