Credits
Exactly what you pay for when the agent works
Artificial intelligence is not billed by the hour or by the user. It is billed for what it reads and what it writes. This page explains that arithmetic without jargon, with figures measured on this very system.
Where it starts
What a token is
A token is a small piece of text, roughly three quarters of a word. "Lease agreement" is about four tokens. A written page runs around 450. It is the unit language models measure work in, the way a print shop measures in sheets rather than in ideas.
Why you pay for what it reads and what it writes
A model does two things and both of them cost: read the text that arrives and write the text it returns. There is no hourly rate, no charge per person connected, no charge per document stored. That is why somebody asking for a summary of two pages consumes far less than somebody asking to reorganize a fifty-page manual, even though both took the same time to type the request.
Ofivia translates that arithmetic into credits so you never have to think in tokens. One credit is US$0.001 of real model spend, so a thousand credits is a dollar. The number recorded is not an estimate of ours: it is the cost the engine reported for that turn, and when the engine reports none it is computed from our price table and flagged as such.
And what is not billed
Uploading documents, indexing them, searching them, walking the link graph and recalling what was already discussed consume no credit. That runs on your own server: indexing with a local model, search as a query against your database. Messages between people do not count either. Credit is consumed when the agent works, and only then.
Anatomy of a turn
Why one turn weighs eighty thousand tokens
When you write to the agent, your sentence does not travel alone.
-
New text you wrote 1.2%
Your question and whatever changed since the previous turn.
989 tokens per turn
-
Text the agent wrote 1.4%
The answer, plus tool calls and file edits.
1,140 tokens per turn
-
Context re-read from cache 83.2%
The department brain note, its memory, open tasks and the history. It rides on every turn, which is why it is cached, and why a turn costs cents rather than dollars.
69,949 tokens per turn
-
Context written to cache 14.2%
The first time that context enters the conversation.
11,946 tokens per turn
Along with your question travels the department's main note, with its description, its memory block and its open tasks. So does the summary of the parent department and of the connected departments. So does the history of what has already been said, and the list of files that chat can reach.
That is why the agent knows what you are talking about without being briefed every morning, and also why a turn weighs eighty thousand tokens instead of two hundred. Almost all of that weight is context being reread rather than new text, which is why it is cached: rereading costs far less than sending the whole thing again.
US$0.19 was the average cost of an agent turn on this system, measured over 119 real turns.
Repricing those same tokens against our internal table would give US$17.03 instead of the US$22.48 the engine reported. We store the engine's number, the larger of the two, because it matches the spend that actually happened.
The table that prevents a bad budget
A million tokens, in two different units
The same figure is two thousand pages or twelve turns depending on what you use it for. Budget with the second one.
| One million tokens is | Amount |
|---|---|
| Characters of text | 4,000,000 |
| A4 pages of 1,800 characters | 2,222 |
| One-page notes | 2,222 |
| Five-page documents | 444 |
| Twenty-page reports | 111 |
| Agent turns, at 84,024 tokens per turn | 12 |
Where you see it
Settings → AI usage
There is no report to request from us. The screen is inside the product and any company owner or administrator can open it.
Budget this month
US$40.00
Spent
US$21.60
| Day | Process | Turns | Credits |
|---|---|---|---|
| Jul 24 | agent.chat | 18 | 3,420 |
| Jul 23 | agent.scheduled | 6 | 1,140 |
| Jul 22 | agent.chat | 11 | 2,090 |
| Jul 19 | agent.chat | 9 | 1,710 |
- Budget and spend. How much credit is assigned and how much the company has used this cycle.
- Fourteen days, day by day. The detail by date, by model and by process, meaning which part of the system made each call.
- Downloadable monthly log. One line per call, stored on disk per company. It records the size of the text covered and never the content.
- Request more credit. The owner asks for a top-up from the same screen, and it is recorded with who approved or rejected it.
Today the system records and displays usage, but it does not stop the agent when the budget is reached. If you go over, you will see it on this screen and we will tell you. We say that here rather than promise an automatic cutoff that is not yet wired into the agent path.
Accounting
Why a retry cannot charge you twice
The boring half of billing is the half that decides whether you trust the invoice.
-
One key per generation
Every call to the model carries a unique generation id, and the accounting row is stored with that id as its primary key. Writing the same call twice inserts zero rows.
-
One transaction
The spend increment, the audit line and the daily rollup go in together or none of them goes in. There is no in-between state where you were charged but nothing recorded it.
-
Retries without fear
If the network drops or the database is congested, the call retries up to six times with growing backoff. A retry cannot charge you twice, and a burst cannot lose a charge.
It is covered by a parallel concurrency test with twelve assertions: no lost updates, no race conditions in the idempotency key, and no bleed between companies. Go back to the plans whenever you want to see which allotment belongs to each one.
Start free. Pay when it earns it.
Starter credit so you can try it against your own real documentation. No card and no sales call. If it works, you pick a plan.
- Starter credit on the house
- No credit card
- Every feature included