What does an AI voice agent cost? An honest breakdown

The real per-minute floor of a production voice agent, why some platform fees are mostly markup, and how to read a quote so you know what you’re actually paying for.

Pricing for AI voice agents ranges from a few cents a minute to six-figure annual contracts, and the spread has less to do with quality than with how each vendor chooses to charge. Voice quality is nearly a solved problem now; the honest way to read any quote is to start from what a call actually costs to run, then ask what the rest of the bill is buying.

The commodity floor

A production phone call is assembled from speech-to-text, a language model, text-to-speech, telephony, and hosting. Together those run about 4–5¢ per minute, so a five-minute call costs roughly $0.21–$0.26. Web-based calls, which skip carrier telephony, land closer to 3¢ per minute. Those numbers are the floor. No vendor escapes them, and no vendor is far below them, because everyone draws from the same commodity components.

How the named platforms price

The published shapes tell you where the margin sits. Read them as structures, not final invoices — your usage still passes through underneath.

• Vapi: roughly a $0.05/min platform fee plus pass-through of the underlying models and telephony
• Retell: around $0.055/min in a similar shape
• Bland: roughly $0.11–$0.14/min all-in
• PolyAI: enterprise, reportedly from about $150k/year

When a per-minute rate sits well above the 4–5¢ floor and includes little else, you are largely paying a markup on the same stack you could assemble yourself. That is fine if the platform earns it elsewhere — but you should know that is what the number is.

What actually justifies a fee above the floor

The commodity is the call. The product is everything around it: an operator console where you can see what happened, fix what is wrong, and prove an outcome landed. That is where real cost — and real value — lives, and it is the part a bare per-minute meter never prices.

• A console built for the operator who owns the phone line, not the engineer who built the bot
• Write-back to your system of record with a verifiable id, so an outcome is provable
• Warm transfer to a human, with a report when the human didn’t answer
• A compliance ledger: consent, disclosure, recording, and cross-channel frequency
• Self-serve edits — change it, test it, ship it, no vendor ticket

How to read a quote

Separate the meter from the platform. Confirm the per-minute number is near the commodity floor, then ask what the platform fee delivers beyond minutes. Per-minute-only pricing surrenders the value of the outcome to a commodity meter; a hybrid model — platform fee plus usage — prices the product and the call separately, which is the only way to tell whether you are buying software or buying markup.

The call is a commodity at 4–5¢/min; the console is the product. A fair quote shows both — a near-floor usage rate plus a platform fee you can trace to real capability, never a marked-up meter dressed up as a platform.