OpenClaw's bill is dominated by background traffic: heartbeats, scheduled skills, and monitoring run around the clock even when you're not talking to it. Routing that background chatter to a cheap model (or a flat-rate endpoint) typically cuts the bill more than anything you do to the interactive sessions.
The background loop doesn't need frontier quality. It needs to notice when something requires attention. Configure the routine path to a budget model and let escalation call the big model only when a skill actually engages.
Every scheduled skill is a recurring API cost. List them, kill the ones you stopped caring about, and stretch intervals (a 5-minute check that could be hourly is a 12x cost difference on that skill).
System prompt, skill definitions, and memory preamble are nearly identical on every call, which is exactly what prompt caching discounts. On providers with cache-aware limits this also multiplies your effective throughput.
A 24/7 agent is the textbook case where per-token billing loses: the usage never stops. A flat monthly price converts OpenClaw's whole cost profile into a constant.
The provider-agnostic tactics (prompt caching, retry budgets, batch APIs) are in the general playbook.
Because OpenClaw is never idle: heartbeats, scheduled skills, and monitoring run continuously in the background. Checking your provider's usage log usually shows a steady baseline of small calls all night. That baseline, not your conversations, is most of the bill.
Only momentarily. The background schedule resumes. The durable fixes are pruning scheduled skills, routing background calls to a cheap model, and capping the whole thing with flat-rate pricing.
The structural version of all of this: run OpenClaw on a flat monthly price, and the bill stops being a variable to manage. 2-minute OpenClaw setup → · Best models for OpenClaw →