Spend credits well

Your workspace shares one pot of AI credits. Your plan adds credits for every seat each month. Those refill on your billing date and don’t roll over, so they’re spent first. Credits you buy never expire, and you can top up by hand or let it happen on its own.

The first month free has no monthly credits. It runs on your welcome credits and the 1,000 each new teammate brings. On free credits alone, the priciest models wait until you buy credits or pick a plan.

The trick is simple: see which jobs use credits, and check whether what comes back is worth it. Current prices are with the plans on the homepage.

Look at usage first

Settings → Usage shows the last 30 days of spending. Filter by feature and open any row to see who spent it and on which models. It also shows how much prompt caching saved you.

Then compare the cost with what you got. A long research job might be worth every credit. A routine that runs every hour and nobody reads probably isn’t.

Coding runs and app builds

These use more credits than chat. Started from the desktop app, they can run on a coding agent already on your computer instead, on your own subscription, for no credits.

You can make that the rule. In Settings → Machines → Spend, an admin can switch off credit runs for machine tasks and for app builds, separately. Those then run only on someone’s own computer, and a browser says plainly that it can’t start one. Everything else, like chat, routines, voice and summaries, carries on as normal.

Spend

Price ceiling Highest model price, in dollars per million output tokens. Leave blank for no ceiling. No ceiling

Turns per run How many steps one Machine run may take before it writes up what it has, from 1 to 100. 25

Machine tasks on credits Off: Machine tasks run only on a member's own machine, from the desktop app.

Apps on credits Off: apps are built only on a member's own machine, from the desktop app. Apps the team has already built keep thinking on its credits.

Set a price ceiling

Also in Settings → Machines → Spend: a highest model price, in dollars per million output tokens. Every model picker in the workspace then only shows models under it.

Nothing breaks when you set one. A machine someone pinned to a pricier model just answers on a cheaper one your team allows. A cheaper model may answer a little differently, so check a few results afterwards. Leave it blank for no ceiling, which is how every workspace starts.

Choose what machines read

Machines read their context again on every step, so what they read is the biggest lever on your bill. Settings → Context shows exactly what goes into every request, biggest first.

Chat. By default, machines also read earlier machine replies, and long replies add up. Turn off Include Machine messages to leave them out. Machines still see the thread they’re in and can read recent chat when they need to.

Docs. Each doc has a Machine context setting. Title only sends just the title, and the machine opens the page when it needs it. That’s the default. Always in context sends the whole page with every request. Hidden from Machines keeps the page away from machines entirely, while people can still read it.

Roadmap P-3

How we deploy P-19

Retro notes, March P-24

Hiring plan P-31

Machine context

Always in context Always seen by Machines.

Title only Machines see the title, and the contents on request.

Hidden from Machines Never seen by Machines.

Keep the always-on list short. A long doc marked Always in context is paid for on every step of every run, not just once. You can change any doc’s setting right from Settings → Context.

Files. The same setting sits on every file, with Name only as the default: machines see what is in Files by name and open a deck or a sheet when a question needs it. Always in context sends a file’s text with every request, so save it for the one deck everything leans on.

Tasks and the grey zone. Open tasks bring their descriptions and comments. Finished ones send only their title and status. After 14 days they move to the grey zone with older chat and calls, which machines skip by default. That’s how they stay sharp and affordable on day 400.

Limit the steps in a run

A machine works in steps: read something, do something, read the next thing. Each step reads everything gathered so far again, so long runs are the pricey ones. Turns per run, on the Spend page, caps how many steps one run can take before it writes up what it has. It’s 25 unless you change it, anywhere from 1 to 100. A run that hits the limit says so, and mentioning the machine again picks up where it stopped.

A few more caps live there too. Daily image cap limits what the team spends on pictures in a day, and 0 turns pictures off. Credits per job limits one job where a machine shares work with your other machines: 1,000 unless you change it. And conflict checks can be switched off if your team doesn’t need them.

Free habits

• Give routines the smallest context that works. A short summary is usually enough, and the routine’s machine trims it when you ask. The whole workspace read every hour is the priciest thing a team can leave running. See Set up routines.
• Clear an app’s thread when it gets long. A long thread is paid for on every turn.
• Turn effort up only where it pays. A triager filing tickets is fine on low effort. A reviewer arguing about architecture deserves more.
• Ask once, clearly. Three vague mentions cost more than one that says what you want back. Each mention is its own run, billed on its own, even one sent while the machine is still busy.

Billing and top-ups

Everyone can see the balance, the plan and the charges in Settings → Billing. Changing them takes an admin or, on Grow, someone an admin gave billing access.

When credits run out, machines pause until your plan’s credits refill. To keep going sooner, buy more, or turn on automatic top-up and choose the balance that triggers it. Top-ups follow your balance, not the calendar, so a quiet month costs nothing extra.