The long-call trap

Here is a puzzle from a real month, numbers rounded and the company anonymized. A voice-agent operator had a client running reliably at about $350 a month. One month the bill came in at roughly $700 — double — with no new customers, no new features, and no change in total call volume worth noticing. The cause took real digging to find: automated callers had started dialing the client's voice agent, and calls that normally ran three or four minutes were stretching to ten.

Same product. Similar minutes. Twice the cost.

Per-minute thinking says that's impossible. The mechanism that makes it not only possible but inevitable is the subject of this chapter, and it is the single most important thing to understand about conversational AI cost.

The mechanism: you pay for the conversation so far, every turn

A conversational model has no memory between requests. To respond to the ninth minute of a call, it must be handed the transcript of the first eight — the system prompt, everything the caller said, everything the agent said back. Every turn re-sends the accumulated context, and input tokens are billed every time they're processed.

So the cost of a call's next minute is not constant. It grows with everything that came before it. Minute nine processes a context several times larger than minute one did — which means minute nine costs several times more than minute one, for the same sixty seconds of talking.

Put formally: per-minute cost rises roughly linearly through a session, which makes total session cost grow superlinearly with session length. Here's the shape, with deliberately simple illustrative numbers:

(Illustrative: a 2,000-token system prompt growing ~1,000 tokens of transcript per minute. Your constants differ; the shape doesn't.)

Sum it and the trap closes: on these numbers, one 9-minute call costs $0.27 of LLM inference, while three 3-minute calls — the same nine minutes of talking — cost $0.135. One long call is twice the cost of the same minutes split into short calls.

That is why the operator's bill doubled while the minutes barely moved. The automated callers didn't add volume. They added length, and length is where the cost lives.

What per-minute math gets wrong

Most of the voice stack bills, budgets, and prices in minutes — it's the unit everyone can see. But the mechanism above breaks minute-math in three specific ways.

Averages hide the tail. "Our average call costs $0.04 per minute" is true and useless when a P95 call costs several times the P50. The expensive calls are exactly the ones an average is designed to smooth away.

Flat surcharge multipliers mis-state cost in both directions. Some platforms acknowledge the problem with a step function — past some context size, minutes bill at 1.4× or 1.5×. It's better than nothing and still wrong twice: short-to-medium calls subsidize the multiplier they never trigger, and genuinely long calls cost more than the multiplier admits. Operators who reconcile invoices against expectations report landing over, not under.

Per-minute pricing quietly subsidizes long sessions. If you sell in minutes — bundles, per-minute rates — your longest-session customers are your most under-priced, and nothing in your pricing tells you. The trap compounds: the customers who feel like your heaviest users are also your least profitable per minute delivered.

The unit that survives: the call

The honest unit for conversational AI is not the minute. It is the session — the call, the chat, the agent run — priced by its distribution, not its average.

Three numbers tell you what a per-minute average cannot:

• Cost per call at P50, P85, and P99. The gap between P50 and P99 is the size of your trap. A narrow gap means minute-math is a tolerable lie; a wide gap means it's an expensive one.
• The cost curve within a call. Where does the marginal minute get expensive for your prompt size and turn rate? That number sets your cap (chapter 5).
• Anomalies against the curve. "This call cost 3× the norm for its length" is an alert worth waking someone for — it's how you catch the automated-caller incident in week one instead of on the invoice.

This isn't just voice

The mechanism is context growth, not audio. Long chat sessions accumulate transcript the same way. Agent runs accumulate tool calls, retrieved documents, and intermediate reasoning — often far faster than a human conversation does. Anywhere a session carries its history forward, cost per step rises through the session, and the long tail of sessions carries a disproportionate share of the bill.

What to instrument

If you take one action from this chapter: stop logging minutes alone. Per session, record —

1. Session length (minutes, turns, or steps — whatever your unit of progress is).
2. Context size per turn (input tokens actually processed, not just generated).
3. Cost per session, kept as a distribution: P50 / P85 / P99, tracked weekly.

With those three, the trap becomes a curve you can see, a cap you can set, and an anomaly you can catch — before the invoice does it for you.