Most voice AI operators pick a pricing model by copying whoever they last watched on YouTube. That's how you end up with per-minute pricing on a deal that should have been a flat retainer — and a client who churns the first month they get a slow lead month.
There are only three models that matter. Here's the real math on each, when it wins, and the trap hiding in the most popular one.
Model 1: Per-minute (usage-based)
You charge the client a markup on call minutes. Your cost is the stack: Vapi/Bland platform fee + the LLM + STT/TTS + telephony.
The real cost floor in 2026, all-in, lands around $0.07–$0.13 per minute depending on whether you're on Vapi (you pay for the orchestration plus each vendor), Bland (more bundled), or Retell (in between). Add your margin and you're quoting the client somewhere around $0.20–$0.35/minute.
Run the math on a real account: an inbound receptionist for a home-services company taking 1,200 calls/month at an average 3.5 minutes = 4,200 minutes. At $0.25/min that's $1,050/month. Your cost is ~$420. Gross margin ~$630.
When per-minute wins: high, predictable volume where the client already thinks in cost-per-call (call centers, lead-gen, anything replacing human phone labor they currently meter).
The trap: your revenue is now hostage to their volume. A slow month for them is a slow month for you. Worse, you've trained the client to watch a meter — every invoice is an invitation to question usage. And you've capped yourself: you can't make more than your markup × their minutes, no matter how much value the agent delivers. Per-minute pricing turns a software margin into a utility margin.
Model 2: Flat monthly retainer
You charge a fixed monthly fee for the deployment, regardless of volume. You eat the usage cost.
Same home-services account: you quote $1,500/month flat. Your usage cost is ~$420, plus say $80 amortized for maintenance. Margin ~$1,000 (67%). The client gets a predictable line item and stops thinking about minutes entirely.
When flat wins: almost always, for SMB clients. It's the easiest to sell ("one number, no surprises"), the easiest to forecast on your side, and it decouples your revenue from their call volume. It also lets you price on value (what's it worth to never miss an after-hours lead?) instead of cost.
The risk: a volume spike eats your margin. Mitigate it with a fair-use ceiling — "up to 5,000 minutes/month included, $0.10/min after." You almost never invoke it, but it caps your downside and makes the proposal look considered.
Model 3: Hybrid (base + usage)
A flat base that covers your fixed cost and a healthy margin, plus a smaller per-minute rate above an included bucket.
Example: $900/month base including 3,000 minutes, then $0.12/min. On the same account (4,200 min), that's $900 + (1,200 × $0.12) = $1,044. Margin similar to per-minute, but with a floor.
When hybrid wins: larger accounts with genuinely variable volume, or when the client insists on usage-based and you want a floor under it. It's the most defensible model for a $5k+/month deal because it mirrors how the client already buys infrastructure.
The decision, in one line
- SMB, one use-case, you want clean margin and easy renewals → flat retainer.
- Enterprise-ish, high predictable volume, buyer thinks in cost-per-call → per-minute (or hybrid).
- Variable volume or a buyer who demands usage pricing → hybrid with a base.
Two pricing mistakes that quietly kill agencies
1. Pricing the build like an afterthought. Charge a real setup fee — $1,500–$5,000 — separate from the monthly. It filters tire-kickers, funds your implementation time, and anchors the monthly as the "cheap" part. Free builds attract clients who churn.
2. Never raising prices on existing clients. Your first ten clients are your cheapest forever if you let them be. Build a 10–15% annual escalator into the contract from day one. Nobody churns over it; everybody would have negotiated it away if you'd asked later.
The model matters less than the discipline. Pick one, write it on the proposal as a single confident number, and stop letting the client price your work by the minute.