Dedicated capacity

On plans with dedicated capacity, the work runs on a separate machine brought up for your organisation rather than on the shared service. The price has two parts: tokens and machine time.

Model requests go only to this machine: nothing is sent to third-party model providers. If the machine is not ready yet, the work waits for it rather than going to a provider.

The plans themselves and how they differ from the others are described on the Plans page.

The model determines the machine

On these plans you choose which model to run. That determines which machine is brought up — and so each class has its own hourly price. The classes and prices are on the pricing page.

You meet the choice twice, and it is one and the same choice. The first time is right when you switch to the plan: the change window does not close until a model is named, because the plan does not work without one. The second is on the dedicated capacity card in “Settings → Billing”, where you can change it later.

The classes differ in the tasks they suit: the basic one is faster and cheaper, the advanced one is stronger on complex tasks. The basic one has a peculiarity worth knowing in advance: text artefacts — tasks, formulations, research and conversations — sometimes contain Latin script, Chinese characters and strange words. This does not affect the quality of the work.

Two modes

Dynamic. The machine comes up by itself when there is work and shuts down after more than 20 minutes idle. Suitable if tasks do not come in all the time: you do not pay for the time the machine is off.

Permanent. The machine stays up until you switch it off. Suitable when work is steady and you do not want to wait for preparation every time.

The mode is shown on the dedicated capacity card in “Settings → Billing”.

The model is chosen before the machine starts: while the capacity is running it cannot be changed — the machine itself depends on the model.

Preparation takes time

The machine has to be prepared: the environment deployed and the model loaded — that is tens of gigabytes. Usually 10–20 minutes.

All this time the card shows that the machine is being prepared and how long preparation has been going on. There is no need to restart tasks — they start on their own as soon as the machine is ready.

One machine per organisation

However many people press “Switch on”, one machine comes up. The second person and everyone after them are shown the very same one — it is not separate capacity for each person but the organisation’s shared machine.

You can switch it off when nothing is running

The “Switch off” button is available, but it only works when the machine is idle: while tasks, “Ask” conversations or research are running, the app refuses and says what exactly is running right now. This way a task is not cut off halfway because someone pressed a button.

Automatic shutdowns — on idle time and on balance — work differently: nobody pressed anything there, so the machine first lets whatever is running finish and only then shuts down.

While the machine is shutting down it has no button — the state is already stated, and there is nothing to press.

Balance: we switch off in advance

The capacity is switched off before the balance runs out, not at the moment it hits zero: when what is left would last less than an hour of work at the current rate of spending. The card warns in advance roughly how long the balance will last at the current rate — spending is counted across both items at once, tokens and machine time.

Task execution stops at that point, the machine stops taking new work and shuts down; after you top up the balance, the task can be resumed from where it stopped.

If the machine does not respond

We constantly check whether the model responds and restart it ourselves on failure. If the machine does not come back, we take another one and bring the work up on it. On the card this shows as preparation; nothing is required from you.

What is not here

Dedicated capacity does not expose a single port to the outside: the model listens only on the machine’s internal address, and the app connects to it itself, with an outbound connection. Nobody can connect to the machine from outside — neither you nor anyone else.