Usage

Every plan comes with a monthly allowance for metered usage. This page explains what spends it, what doesn't, and how to keep an eye on it. For your current allowance and where it's going, see Settings → Billing.

On a paid plan, Auto on a CPU sandbox is the free floor: the allowance bar does not move at all. It fills only while metered work runs — here, the two cells a GPU is attached for — and stops the moment the GPU comes off.

What's metered

Three things draw down your allowance:

• Models: calls to the metered models, which is everything above Auto. More capable models and higher reasoning effort cost more, because they do more work per call.
• Sandbox: time on a metered runtime, charged by the minute. A GPU costs more than CPU, and a bigger GPU costs more than a smaller one. The meter runs while a paid runtime is attached to your work, and an idle sandbox shuts itself down so it doesn't keep charging.
• Web tools: the agent's web-research calls, when it searches the web, fetches a page, or extracts structured content from one. These run against a paid provider, so they're metered per call.

Editing, branching, running cells, and reading output are not metered.

The free floor

On a paid plan, the Auto model and CPU runtime are never metered. Run Auto on a CPU sandbox all day and your allowance doesn't move. That's what the picture above is showing: the bar sits flat while real work happens, and only starts filling when a GPU attaches.

That's deliberate. The everyday loop, where you describe a task and let Auto write and run it on CPU, stays free on a paid plan. Your allowance comes into play when you reach for something heavier, like a paid model or a GPU.

On the free tier, the rule is different

The free tier has no cheaper floor to fall back to, so Auto and CPU are exactly what its allowance buys. Every run counts against it, including plain Auto-on-CPU work. If you're on Free and you expected the meter to stay still, it won't.

What this means in practice:

• Nothing is blocked. Free runs are never stopped or paused, whatever the meter says. There's no premium model or GPU to cut off, so there's nothing to gate.
• The expensive things aren't reachable anyway. The model picker is locked to Auto, and every GPU needs a paid plan. So beyond Auto and CPU themselves, the only other thing a free account can spend its allowance on is a web-tool call. Paid models and GPU minutes, which are most of a paid bill, aren't in reach.
• Frontier models on Free are still possible, but not through the meter. Two routes: bring your own key, which bills your Anthropic or OpenAI account directly, and a paired local runner, which runs the turn on your machine against your own CLI subscription. Neither touches your Clusy allowance.

Your allowance

The allowance resets each billing cycle. On a paid plan, when it runs out, metered work stops and you drop back to the free Auto-on-CPU loop, unless you've turned on pay-as-you-go.

A run that started inside your allowance is never cut off partway through. There is no per-run limit: within your allowance a run is uncapped, on purpose, so an hour-long agentic GPU run isn't killed mid-flight. The allowance is checked before each run, so the bound on an overrun is the next run, which degrades to Auto on CPU and tells you why.

Reading the usage meter

The usage meter in Settings → Billing shows where your allowance is going:

• how much of it you've used this cycle, as a percentage,
• a breakdown by category across Models, Sandbox, and Web tools,
• which models and which runtimes took the most,
• and a projection of how many days are left at your current rate.

If the projection says you'll run out early, lean on Auto and CPU more, dial back effort, or look at pay-as-you-go.