01
Input and output are priced apart.
Everything you send counts: the question, the instructions around it, the conversation so far, and any passages retrieved from your documents. What comes back is charged at a higher rate per token.
Nothing on this page is a quote. It is the arithmetic, the failure modes and the controls, so you can work out your own number.
The compute bill is yours, so you can always see it.
Every AI vendor has a compute cost. Most of them are paid to keep it away from you. Ours arrives on an invoice with your name on it, from a provider we are not a party to, at their published list price. That is uncomfortable in one specific way, which this page is about: a bill you can read is a bill you can be surprised by.
Priced by the token, by the model.
INFERENCE · HOW IT IS PRICED
Inference is metered in tokens, which are fragments of text rather than words. Your provider publishes a price per million tokens for each model, and charges input and output at different rates, because generating text costs them more than reading it. A request costs what its tokens cost at the list price of the model it ran on, and nothing else.
There is no per-user component, no platform fee riding on top and no rounding in our favour, because the invoice is not ours. We never see it unless you show it to us.
Charged by your provider
Driven by your usage
Never touched by us
02
Each model has its own price.
The catalogue we serve holds 104 active models from 18 labs. A frontier model and a small one differ by more than an order of magnitude per token, and the routine work usually does not need the frontier.
03
The price is the provider's list price.
We do not resell compute, so there is no wholesale rate we are protecting and no margin to defend. If the provider changes their price, your bill changes on their schedule, not on ours.
04
Storage and infrastructure are separate.
The platform itself has a running cost on your bill: storage, the index, compute for the dashboard. It is small next to a busy month of inference and it is listed further down this page.
Show the arithmetic.
WORKED EXAMPLE · ONE QUESTION
Estimate, not a quote
A knowledge question, priced end to end.
Take a typical question asked against your own documents. The request carries the instruction template, the passages retrieval pulled back and the question itself: on the order of 10,000 input tokens. The answer runs to about 1,000 tokens out. At a mid-tier frontier model's list price, the whole exchange lands at about four to five cents.
- 01It is an estimate and it is labelled one. Token counts vary with the question, the number of passages and the length of the conversation. Provider prices change.
- 02The model choice moves it more than anything else. The same question on a smaller model costs a fraction of this, which is why routine work does not belong on the most expensive thing in the catalogue.
- 03Multiply by reality, not by hope. A few thousand questions a month at this shape is tens of dollars, not thousands. A team pasting whole documents into long sessions is a different arithmetic entirely.
| Line | Tokens | List price | Estimated cost |
|---|---|---|---|
| Input: template, retrieved passages, question | 10,000 | Order of $3 per million | about $0.030 |
| Output: the answer, with citations | 1,000 | Order of $15 per million | about $0.015 |
| One question, end to end | 11,000 | Mid-tier frontier model | ✓about $0.045 |
Illustrative estimate, not a quote and not a measured average. It assumes one mid-tier frontier model at the order of magnitude its list price sits at today, and it excludes the platform infrastructure line further down this page. Your provider's published price list is the only authority on what you will actually pay.
Four ways a bill jumps.
COST · WHAT MAKES IT MOVE
A
Somebody pastes a long document.
A hundred-page contract dropped into a chat box is a hundred thousand input tokens, and it is charged again on every subsequent turn of that conversation. This is the single most common cause of a surprising week, and it is usually one person doing something entirely reasonable.
B
Long multi-turn sessions.
Conversation history is resent with each turn, so a thread that runs all afternoon charges its own transcript repeatedly. A fresh session for a new task is cheaper than a long one, and better answers usually come with it.
C
A volume step-change.
A department that starts using the platform properly, or a module that begins calling a model in a loop, moves the bill by a multiple rather than a percentage. This is the good kind of spike, and it should still be a decision rather than a discovery.
D
The provider changes their price.
Model prices move, in both directions, and new models arrive at new rates. Because we do not sit between you and the bill, a change reaches you directly. We refresh the catalogue as providers change it and flag what moved.
A rising query count with flat value is a cost problem, not a success.The measure we hold ourselves to
Four things that hold the ceiling.
CONTROLS · BOUNDED BY DESIGN
Enforced, not advised
A quota that fails closed is a budget you can actually keep.
Cost control by policy document does not survive contact with a busy week. These four controls are in the deployment, and the first of them refuses rather than warns.
- 01Per-person token quotas that fail closed. Each call reserves against the person's quota before it runs and settles against the provider's own record of the call afterwards. If the meter cannot be read, the request is refused rather than allowed through unmetered.
- 02Usage by person, visible to that person. Everyone signed in can see their own consumption; administrators can see it by person and by day. It is reported in tokens, because the product holds no currency at all.
- 03A budget alarm on the account. Set against a number you choose, watched as part of managed operations, and raised with you rather than absorbed quietly.
- 04A smaller model for routine work. Model choice is configuration inside the same governed route, so moving ordinary traffic off the most expensive model in the catalogue is a parameter change, not a project.
metering · reserve-then-settle
FAIL CLOSED
inference request → governed route
[meter] quota resolved from role grant
Reservation opened✓ within quota
Content policy attached✓ pinned version
Settled from provider record11,042 tokens
[meter] next caller, quota exhausted
Reservation refused✕ no tokens remaining
Meter unreadable✕ refused, not allowed
A quota that cannot be checked is a quota that denies
Illustrative outputtokens only
Metering·Settled against the provider's own record, home region
Metering, precisely
Metering covers the home region. A call routed to another region is logged and not metered, and under default global routing that happens, so per-person usage can under-count. A failed settlement deliberately overstates rather than understates, which is the direction we would rather be wrong in. Pinning a deployment to one jurisdiction narrows this, at a real capability cost set out on data residency.
What the platform itself costs to run.
PLATFORM · BY POSTURE TIER
Inference is the variable. The platform underneath it has a steady monthly cost on your bill, and that cost is driven almost entirely by how hardened you asked the deployment to be. Private networking is the expensive control, and it is worth its price for the reasons on the posture tiers page, but nobody should meet the number for the first time on an invoice.
| Posture tier | What it adds to the infrastructure | What we have seen, monthly |
|---|---|---|
| Baseline | Single-account isolation, encryption in transit and at rest, mandatory content policy, least-privilege identity policy, everything as code. No private networking, no account audit trail. | Roughly $5 to $10 |
| Standard | Adds a private network endpoint to the managed model service, network-attached chat compute, account-level audit logging with file validation and one-year retention, enforced MFA, provider-managed keys. | About $110 |
| Strict | Everything in Standard, plus customer-managed encryption keys with rotation and a residency pin that refuses in words rather than rerouting. | About $110, plus key management |
| All in, light team usage | One real deployment, infrastructure and inference together, on a light workload. | ✓An estimated $40 to $120 |
Figures observed on deployments we operate, excluding inference except in the last row, and rounded. They are not a quote and not a commitment. A security review will usually want Standard, and the gap between the first two rows is the real cost of that decision.
The finance questions.
QUESTIONS · COMPUTE
01Why do you not just bundle compute into a monthly fee?
Because a bundled number has to include a buffer for the month you use more than expected, and that buffer is margin you cannot see. The moment we sit between you and the provider, we are pricing your usage risk and taking a spread on it. Keeping the invoice in your account removes both, and it means the number you are optimising is the real one.
It has a cost to us in comfort: a variable bill is harder to sell than a flat one. We would rather explain the variable than hide it.
02How do we forecast this before we deploy?
In the assessment, from your own shape: how many people, doing what, against how much material. Multiply an estimate like the one above by realistic volumes, add the platform line for the tier your security review will accept, and set a budget alarm at the number that would concern you. Then check it against the first two months rather than trusting the model.
What we will not do is give you a per-user figure. There is no seat concept in the product, so a per-user number would be an invention dressed as a forecast.
03Does a bigger knowledge base cost more per question?
Storing the documents costs a little, and it is part of the platform line rather than the inference line. What moves the per-question cost is how many passages get pulled into the request, not how many documents exist behind them. A larger corpus mostly changes retrieval quality, not token volume.
The limits of the knowledge base itself, including the fact that everything loaded into a deployment is readable by every user of that deployment, are stated on chat and knowledge.
04Who watches the bill month to month?
You own it and can read it without asking anyone. Cost watch is part of managed operations: the budget alarm, a look at what moved, and a note in the operations journal. Every caller runs under its own tagged profile, so spend separates in your cost reporting by model and by consumer rather than arriving as one undifferentiated line.
05Is there ever a case where inference is the wrong answer on cost grounds?
Often. A deterministic tool that costs nothing per run beats a model that costs cents per run and sometimes gets it wrong, and most of the modules organizations actually use day to day never call a model at all. Where a lookup, a rule or a calculation will do the job, that is what should be built. Where AI does not belong is the longer version of that argument.
Why there is no markup
No markup is not a promise we make. It is a shape we cannot get out of.
The platform runs in an account you opened, so the provider invoices you. We are not a party to that transaction, which means there is no place in the architecture for a margin to live, whatever we might say on a web page.
- 01No compute passes through us. We hold no account of yours and resell nothing.
- 02Every caller is tagged. Spend separates by model and consumer in your own cost reporting.
- 03You can check it. The evidence is in your console, not in our invoice.
⎯⎯ Book the strategic assessment ⎯⎯
Your private AI, inside your control ·
A bounded $2,500 engagement that sizes the real cost against your own usage before anything is deployed.