A workspace's own AI spends its AI credit every time it thinks: each step in a
chat, each run of a scheduled job, and the hourly check-in when it is on. Team
includes $100 of AI credit a month, and the free start comes with $3 for two
days. Most of what a workspace spends comes down to four settings: the model, how
often its jobs run, the switches under **Cost savings**, and the account that
pays. Each one is a few clicks, and none of them costs the workspace its memory.

## Pick the model that fits the work

An admin picks the model under **Settings**, then **Model**. The picker groups
each model by what it is built for and shows its price beside the one running
now. Once the workspace has made about twenty calls, the picker also prices the
last 30 days of the workspace's own use on both models before anything changes.

The price of a model is per million tokens, which is hard to picture. The last
column below prices one ordinary step the AI takes (a short answer that re-reads
the conversation so far), measured across GBrain's workspaces:

| Model | Built for | Per million tokens, in / out | One ordinary step |
|---|---|---|---|
| Claude Haiku 5.5 | Quick, simple work | $0.10 / $0.50 | 0.2¢ |
| GPT-6 Luna | Quick, simple work | $0.10 / $0.50 | 0.2¢ |
| Particle-1 | Everyday work | $0.25 / $1.00 | 0.4¢ |
| Claude Sonnet 5.5 | Everyday work | $2 / $10 | 3.4¢ |
| GPT-6.1 Sol | Hard work | $2 / $10 | 2.6¢ |
| Claude Opus 5.5 | Hard work | $4 / $20 | 6.9¢ |
| GPT-6 Astra | The hardest problems | $10 / $50 | 15.4¢ |
| Claude Fable 5.1 | The hardest problems | $10 / $50 | 15.9¢ |

At those prices $100 of credit is about 2,900 steps on Claude Sonnet 5.5 and
about 50,000 on Claude Haiku 5.5.

- **Claude Sonnet 5.5** is where a new Team workspace starts, and it is a good
  choice for a team that talks to the workspace every day.
- **Claude Haiku 5.5** or **GPT-6 Luna** suits a workspace that mostly files,
  sorts and summarizes on a schedule. Haiku costs about a seventeenth of Sonnet
  5.5 per step, and Luna about a twenty-second. They are built for short, high-volume work, so try a long
  report on one before relying on it.
- **Opus, Astra and Fable** are for work that fails on the models above them.
  Try the job on Sonnet 5.5 first.

One model answers for the whole workspace: its chats, and the scheduled jobs too,
unless a skill names a model of its own. A change applies from the next message,
with no restart, and switching back is the same dropdown.

The free start runs on Claude Sonnet 4.6, and its model cannot be changed until
someone keeps the workspace on a plan.

## Pick the account that pays

The same **Model** page has **Which account pays for this**. An admin can sign
the workspace in to their own ChatGPT Plus or Pro plan, and then its OpenAI
models stop spending the workspace's credit. An organization can also add its
own Anthropic or OpenAI key, which that provider bills directly.
[Which AI does what](/docs/workspace/what-runs-where) has the three accounts and
how switching between them works. The free start uses GBrain's accounts only.

The thinking of the AIs people connect, such as Claude, ChatGPT or Cursor, is
paid by those AIs' own plans, not by the workspace's AI credit.

## Run scheduled jobs only as often as someone reads them

Every run of a scheduled job is a paid turn for the workspace's AI. A job that
runs every hour runs about 720 times a month; the same job once a day runs 30
times. So run a job as often as someone reads what it makes: an inbox summary
that is read each morning needs one run before the morning, not one every hour.

**Schedule** lists every job with when it last ran and when it runs next, and an
admin can delete a job there. To change how often a job runs, ask in a chat, for
example: "Move the inbox summary from every hour to once a day at 8:00." Then
check **Schedule** to see the new time.

The morning email at 7:00 is a scheduled job too. Turn it off from the link at
the bottom of any of them.

The jobs GBrain runs for the workspace, such as the overnight upkeep that keeps
the memory tidy, are under **Behind The Scenes Jobs** at the bottom of
**Schedule**. They run once a day and stay on.

## Turn on the switches under Cost savings

An admin finds three switches under **Settings**, in **Cost savings**:

- **Spending brake.** Turn it on and pick an **Hourly limit**, from $2 to $100 an
  hour; it starts at $25. Once the workspace spends that much in any 60 minutes,
  the AI pauses and says so, and it picks up again on its own as the hour passes.
  It is the setting that stops one runaway task from spending a month's credit in
  an afternoon. A workspace that mostly runs a few scheduled jobs does well with
  $2 or $5.
- **Keep long tasks in check.** Turn it on. The AI stops when it repeats the
  same step without progress, and trims old results out of its working memory
  sooner. Very long research may get cut short.
- **Hourly check-in.** Leave it off unless something needs checking every hour.
  Each check that finds something on its list is a paid AI call, so a busy
  list can add up to 24 calls a day; an empty list costs nothing. Scheduled jobs and reminders run on their own times
  either way. A new workspace starts with it off; a workspace made before the
  switch existed may still have it on.

## Start a new chat for a new topic

Each step in a chat re-reads the conversation so far, so a long chat costs more
per message than a short one. When the topic changes, start a new chat. The
workspace still remembers what was saved to its memory.

Web search, page reading and research that come with the plan are paid from the
same credit, and each use shows in the [activity log](/docs/tools/activity)
with its cost.

## See what costs money

- **Billing**, under **Settings**, shows an admin what the workspace can still
  spend: what is left of this month's AI credit plus the organization's credits.
  At $0 the workspace stops answering, and its chats, memory and schedules stay
  put. A workspace on the organization's own key shows what it has spent this
  month with each provider instead.
- **Model** prices the last 30 days of the workspace's use on any model in the
  picker, before switching.
- The organization's **Credits** page adds credits, and can top them up on its
  own when they run low.
- The [activity log](/docs/tools/activity) shows each paid lookup with its cost.

## A low-cost setup in five minutes

1. Under **Settings**, **Cost savings**, turn on **Spending brake** at $5 an hour
   and **Keep long tasks in check**, and check that **Hourly check-in** is off.
2. Under **Settings**, **Model**, keep Claude Sonnet 5.5 for a workspace people
   talk to every day, or pick Claude Haiku 5.5 for one that mostly runs jobs.
3. Open **Schedule** and ask in a chat to move any hourly job to the times
   someone reads it.
4. If an admin has ChatGPT Plus or Pro, sign the workspace in to it on the
   **Model** page and pick an OpenAI model.
5. A week later, open **Billing** and see what is left.