Cost and budgets

An automation’s cost has a simple shape: the machinery is free, and each model-using step costs a small amount every time a run reaches it. A watch that just detects a change and messages you spends nothing at all. A watch that summarizes what changed spends one small call on the runs where something actually changed — and nothing on the quiet days. The cost tracks the thinking, not the vigilance.

Every scheduled check, every change comparison, every deterministic transform, and every delivery is free. That is the deliberate design choice that makes standing automations viable: an automation can be armed for a year, checking a page every hour, and cost effectively nothing across all those thousands of checks — because none of them involve a model unless something happened that needs one.

Every step that uses AI — screening a lead, summarizing, reading an image, generating speech or an image — runs on a model that you chose. The assistant never silently picks one for you. When you describe an automation that needs a judgment step, it presents the fitting options with their costs and lets you decide, so you always know what each run of that step will cost and why.

For each kind of AI work, you get a menu of the models that can do it. Every entry carries what you need to choose well:

The menu is served live, so it always reflects the current lineup. Some entries are resilient options that automatically fall back to an alternate model if their first choice is momentarily unavailable — handy for steps you want to keep working even if one provider has a bad minute. You choose from labels, never cryptic identifiers.

When you create an automation, the confirmation includes a forecast — an estimate like “about $X per month at this cadence” — computed from how often it will run and the per-call estimates of its AI steps. If a step runs inside a loop, its cost multiplies per item, and the assistant says so when it quotes the number, so a “summarize each of the 30 new items” automation doesn’t surprise you. Automations with no AI steps are called out as costing nothing to run.

Before an automation goes live — or any time after — you can preview it: the whole recipe runs with every side effect mocked. Nothing sends, nothing writes to your tools, nothing is charged, and you get back what would have happened at each step. It is the way to prove an automation does what you meant before you trust it with the real thing, and it’s offered right after creation for exactly that reason.

You can set a monthly budget for an automation. As it runs, its AI spend is tallied against that cap, and if it reaches the ceiling the automation pauses itself rather than spending past what you allowed. You’re notified, and resuming it — or raising the cap — is one action. Leave the cap off and the automation runs uncapped. Either way, spend is never a surprise: the cap is a hard stop you set, not a bill you discover.

Every time an automation is created or edited, that version is saved. So there is always a history of how it changed over time — including the fixes it made to itself while self-healing. If you want a fully diffable, external record, an automation’s recipe can also be mirrored to a code repository you own, where each change (and each repair) lands as a commit you can read. Either way, an automation is never a black box that quietly became something other than what you set up. Its evolution is on the record. For how you review and edit an automation, see Building an automation.