Skip to content
Docs navigation: Open models / Reserved capacity

Reserved capacity

RightNow AIOpen modelsReserved capacity Updated

On this page

Reserved capacity holds a number of concurrent requests on one model for your workspace. It is for a team that needs a set number of requests to run at the same time on that model.

How it works

how a reservation works
Text
Illustration of how a reservation works.

A reservation is made in the dashboard, under Capacity. Choose: a model and how many concurrent requests to reserve. Review: the price per hour, what is charged now and what it covers. Pay: it is live at once. Watch: its state, what was charged and what was returned. Change: its end date. The end: on its end date, or when you end it.

You reserve in the dashboard, under Capacity.

What it costs

You pay a price per hour for each reserved concurrent request.

ModelPer hour, for each reserved concurrent request
GLM 5.3USD 1.50
GLM 5.3 FlashUSD 0.75
Ornith 1.5 35BUSD 0.25

A model that is not in the table has no reserved capacity.

Tokens are billed as on any other request, at the list rates in Pricing and credits.

The hourly price is paid ahead from your balance, the same balance that pays for tokens. When you reserve, you pay for the rest of the current hour and the next 3 hours. From then on, 3 hours are always paid ahead.

A reservation is not charged for an hour in which its model is paused.

You choose when it ends, or leave it open. A date you pick ends it at the start of that day, 00:00 UTC. Nothing renews past its end. With no end, it runs while it is paid, until you end it.

Starting and stopping

A reservation is live from the moment you pay.

Live
It is in force, and your requests use it.
Stopping
An owner ended it from the end of its paid time. It serves until then, and no more hours are charged.
Ended
It no longer serves requests.

An owner can end it in two ways:

  • Now. Unused paid time is returned to your balance, except its first 60 minutes.
  • At the end of its paid time. It serves until then, no more hours are charged, and an owner can undo it until then.

An owner can also change its end date. An earlier end date returns the hours that begin after it.

Money is returned to your balance, never to a card.

A reservation ends on its end date, when an owner ends it, for low balance, when a card payment is disputed, or when the workspace is deleted.

Its page in the dashboard shows what was charged and what was returned, by day.

When the balance runs low

A reservation is paid ahead, so it serves to the end of its paid time.

Before you pay, the review shows the amount charged now.

To keep it running, keep your balance above its price per hour. Auto-recharge can do that: set its Balance below amount above the reservation's price per hour. With auto-recharge on, your card may be charged for it.

When your balance does not cover the next hour, the reservation serves to the end of its paid time and then ends. The dashboard then reads Ended for low balance. When recent card funds have reached the daily limit for reserved capacity, it is renewed at the size the funds that can pay for it cover.

Once it has ended, your requests run on shared capacity.

Limits

Who reserves
The account owner, in the dashboard. An API key cannot reserve, change or end a reservation.
What pays
Only credits you bought pay for reserved capacity.
Recent card funds
They pay for reserved capacity within a daily limit. USDC top-ups have none.
Smallest size
2 concurrent requests.
Largest size
The room on sale on that model when you buy. A larger size is refused.
For each model
One reservation.
A request for several answers
Uses one reserved request for each answer.
Large prompts at once
The larger of 2 and a quarter of the reserved number. A prompt is large from the model's large_prompt_threshold_tokens in GET /v1/models.
Tokens each minute
A reservation has its own token limit each minute, set by its size.

In the API

Nothing changes in your code. Every key of the workspace uses the reservation with the same model id.

Requests above the reserved number run on shared capacity at the list rates, or are refused. You choose which when you reserve. A request that does not fit one of these limits follows the same choice.

The answers of POST /v1/chat/completions, POST /v1/responses and POST /v1/messages carry the header X-RightNow-Capacity. It says where the request ran:

reserved; held=N
The request ran on your reservation. N is the number of concurrent requests you reserved.
shared; held=N; spill=1
The request ran on shared capacity, above your reservation.