The question of what AI costs a company almost always gets answered in the wrong currency. The vendor quotes a price per million tokens. The person asking wants to know what next month looks like for twelve people. Between those two sits a chain of conversions nobody likes to do in public, because doing it means showing your assumptions.
So here they are. Every input is stated, and if you disagree with one, swap it and run the numbers yourself.
Where the bill comes from
Five things are metered and nothing else: input tokens, output tokens, web search queries, sandbox seconds when code runs, and storage for documents. There is no per seat charge. Somebody who did not open the assistant this month costs nothing.
That sounds like an accounting detail. In practice it changes the shape of the whole spend. On per seat licensing you pay for sixty accounts of which twenty are actually used. On usage you pay for the twenty, and you stop having the quarterly argument about whose licence to take away.
The rates behind the numbers
| Model class | Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| Efficient | gemma-4-26b-a4b-it | 0.40 EUR | 0.80 EUR |
| Balanced | qwen3.5-397b-a17b | 0.96 EUR | 5.76 EUR |
| Strongest | glm-5.2 | 3.24 EUR | 9.90 EUR |
Add web search at 0.0012 EUR per query, storage at about 0.019 EUR per gigabyte month, and embedding at 0.16 EUR per million tokens when documents are indexed.
Look at the ratio of output to input on the balanced class: six to one. That is the most important number here and I will come back to it.
The usage assumptions
This is normally where guessing starts. These come from measurement on a running platform rather than from a vendor’s calculator:
| Operation | Input tokens | Output tokens |
|---|---|---|
| One grounded answer in a conversation | 1300 | 350 |
| A helper call, for instance rewriting the query | 900 | 250 |
| One retrieved passage in context | 410 | 0 |
| One automation step | 4800 | 1100 |
Why 1300 input tokens for a ten word question? Because what reaches the model is a system instruction, the conversation so far and six document passages. The question itself is a rounding error. Anyone estimating AI cost from the length of the prompt is off by two orders of magnitude.
Twelve people, one month
The profile: twelve people, fifteen messages each per working day, twenty one working days, the balanced model, sixty percent of conversations grounded in the knowledge base, two hundred web searches, three hundred automation runs, twenty gigabytes of documents.
| Line | Per month |
|---|---|
| Conversations | about 27 EUR |
| Retrieved passages in context | about 5 EUR |
| Web search | under 1 EUR |
| Automations, 300 runs | about 3 EUR |
| Storage, 20 GB | about 0.40 EUR |
| Total | about 35 EUR |
Under 3 euro per person per month. For comparison, a per seat licence from a vendor charging by headcount starts around 20 to 30 euro per person per month, used or not. At twelve people that is a difference of roughly 250 to 350 euro a month, call it tenfold.
Before that sounds too good: a team that leans on AI hard changes the shape. Forty messages a day per person instead of fifteen takes the conversation line from 27 to roughly 72 euro. Still cheaper than licences, but the gap stops being tenfold.
What the number excludes
The one off ingest. Twenty gigabytes of text is roughly 430 million tokens to process while indexing, about 69 euro, paid once. If the base grows by two gigabytes a month, add about 7 euro a month.
It also excludes people. Rolling this out, choosing which documents go in, setting permissions and teaching a team to use it costs more than the platform does in the first year. Budget only for licences and the conversation later is unpleasant.
Three levers that actually cut the bill
Shorter answers. Output costs six times what input costs. A system instruction that asks for brevity can take low double digit percentages off the bill with no loss of usefulness.
Fewer passages in context. Six passages at 410 tokens each is 2460 input tokens per answer. Dropping to four saves about a fifth of the conversation line, and you only feel the difference on genuinely broad questions.
Match the model class to the job. Rewriting a note or classifying a ticket does not need the strongest model. At identical token usage the efficient class costs eight times less on input and twelve times less on output. Most of the money companies waste here is not because AI is expensive; it is because everything goes to the most expensive model available.
Checking it against your own numbers
The estimator on the pricing page runs this exact arithmetic and lets you change every variable: people, intensity, model class, knowledge base size, searches and automation runs. The result is an estimate and not a commercial offer, but the inputs are real and come from the same table the billing does.