You pay for what actually ran.
Top up a balance and the platform draws on it as work happens: tokens, searches, runs, storage. No plan to outgrow, no licence per person, no minimum commitment. Every rate is visible in the console before anything runs, and every charge points back to the call that caused it.
What is metered
Each unit is counted where it happens and priced from the catalogue your organization sees.
-
Model usage
Input and output tokens, per model, at the rate published for that model.
-
Web search
Per query, when a conversation or an agent searches the web.
-
Code execution
Per second of sandbox time, counted while your code actually runs.
-
Functions
Per execution of a published function, whether it was called by a URL, an agent or a schedule.
-
Storage and traffic
Stored data per day, bytes sent out, and requests served by object storage.
-
Document store and sites
Per document operation, and per second of processor and memory a site uses to answer.
What would it cost us?
Set the sliders to your own shape of work. The estimate uses the launch rates from the catalogue, states every assumption it makes, and is worth exactly as much as those assumptions.
- Chat and research (messages)
- Knowledge storage (GB)
- Web search (queries)
- Automations (runs)
Estimated monthly spend
per person a month
These are the platform defaults published in the catalogue. Your organization sees its own rates in the console, and nothing is charged until a call actually runs.
The figures are an estimate for guidance only. They are not a quote and not a commercial offer, and the real bill is decided by real usage.
What this assumes
- 21 working days a month.
- An answer averages 1,300 tokens in and 350 out, which is what the platform's own audit trail records for a chat answer.
- Each message also runs two short helper calls, together 900 tokens in and 250 out, for rewriting the query and offering follow ups.
- With a knowledge base in use, 6 messages in 10 carry 6 retrieved passages of 410 tokens, the measured average length of a stored passage.
- An automation run is 2 model steps of 4,800 tokens in and 1,100 out, plus 2 function calls.
- A gigabyte of uploaded documents yields around 21.5 million tokens of text, measured on the documents already in the store, and is embedded once when it is read in.
How you pay
-
A balance, topped up
Add credit when you want to. The balance is the organization's, and every service spends from the same one.
-
Refused before it is spent
A request that would exceed the balance or a limit is refused before it runs, with the reason in the error, rather than appearing on an invoice later.
-
Reserved for the duration of a call
The worst case cost is held while the request is upstream and the real amount is settled when it returns, so concurrent traffic cannot spend the same balance twice.
-
One invoice
Everything in euro, with usage broken down per day, per model and per key.
Guards you can set
-
A limit per organization
A daily or monthly ceiling on everything the organization spends.
-
A budget per API key
An integration gets its own ceiling and is refused at it, so one script cannot spend the department's month.
-
Notification before the wall
The platform tells somebody when a balance is running low rather than only when it stops working.
Questions we are asked before the first call
- Do I need a credit card to start?
- No. You can create an account and look around without one. A card is needed only when you top up a balance.
- Where do I see the rates?
- In the console, per model and per unit, before you spend anything. Your organization sees the catalogue that applies to it.
- Can we bring our own provider account?
- Yes. A deployment can point at a provider you already pay for, and the platform still records usage and cost for your own reconciliation.
- What if a request fails?
- A request that never reached a model is not charged, and money held for a call that failed is released.
Start without a card
Create an account, look at the rates for the models you would use, and decide afterwards.