The dirty secret of AI agents is that the naive version is ruinously expensive. An agent that runs a frontier model continuously, on a always-on VM, reading everything with its most expensive brain, costs more than the employee it was supposed to spare you.
Waveguide’s flat monthly pricing works because of three engineering decisions, and we think you should know what they are — because they’re also why the bill can never surprise you.
1. The right model for the job
Not all agent work needs the same intelligence. Waveguide routes every task to the cheapest model that does it well:
- A fast, inexpensive model reads everything: heartbeat checks, digest triage, sorting incoming signals into act / notify / ignore. This is the majority of wake-ups, and it costs pennies.
- The flagship model does the real work: campaign decisions, lead replies, creative direction, mission planning.
- The heavyweight model is reserved for genuinely hard problems — available on Growth and Scale plans as an escalation, not a default.
The triage layer is the key insight: most signals don’t deserve the expensive model’s attention, and a cheap model is perfectly capable of knowing which ones do.
2. Scale-to-zero compute
Your agent doesn’t idle. Between wakes, its sandbox machine suspends — after sustained inactivity it’s snapshotted and destroyed entirely, then restored on the next signal. You’re paying for work, not for a VM’s heartbeat.
This is also why “always-on” is the honest description: the agent is always reachable by a signal, but almost never running. It wakes in seconds when a lead arrives at 2am; it costs nothing while nobody’s leading.
3. The 80/100/120 ladder
Every plan includes an AI budget ($15, $50, or $200 a month). Every model call is metered against it, to the cent, visible live on your Billing page. Then three automatic guardrails:
- 80% — you’re warned. A heads-up email. No behavior change.
- 100% — the agent economizes. It soft-lands: non-urgent work shifts to the cheaper model, and further usage bills as overage at our cost × 1.3.
- 120% — the agent stops. Hard pause. It waits for you to explicitly resume it. Nothing about this is negotiable by the agent itself.
The worst-case month is a number you can compute in advance, and you chose it. We think of this as the billing equivalent of the approval system: not a promise to behave, but a mechanism that can’t misbehave.
Plans and the comparison table are on the pricing page; the mechanics are in the budgets & billing docs.