What a voice agent costs in 2026 (and why ours costs $0.035 per call)

A 15-minute call on a commercial voice AI API costs between $2.56 and $5.17 in 2026. The same call on the DaCodes voice stack costs about $0.035, measured internally in August 2026. At contact-center volume, that difference decides whether a voice agent has a business case at all.

Where the money goes in a voice call

A voice agent is three models and a lot of plumbing: speech to text, a language model that decides what to say and which tool to call, and text to speech. Add telephony, turn detection, interruption handling and the integration to your system of record. Commercial voice platforms bundle the three models and charge per minute. That is convenient and, at scale, the most expensive line in the project.

The market price

• Commercial voice APIs: $2.56 to $5.17 per 15-minute call, depending on plan, voice and model (DaCodes measurement, Aug 2026).
• At those prices, a contact center handling 100,000 calls a month spends between $256,000 and $517,000 a month on the voice layer alone, before agents, integration or people.

Those numbers are why many voice pilots stop at the pilot. The agent works. The invoice does not.

Our stack and its price

We built the voice layer on open-weights components hosted in the client's cloud: Parakeet for speech recognition, Qwen3 32B as the conversational model, Kokoro for speech synthesis. Orchestration runs on the same runtime as our other agents, with MCP as the tool interface to the CRM, the TMS or the core. The measured cost is about $0.035 per 15-minute call, all three models included (DaCodes internal measurement, Aug 2026).

In a model we ran for a large group in México, the same call volume came out at about $60,000 a year on our stack against about $5 million a year on a commercial voice API (DaCodes, Aug 2026). That is not a rounding difference. It is the difference between a voice agent in every branch and a voice agent in a slide.

What you give up, honestly

• Voice quality: commercial synthesis still sounds a little more natural in English. In Spanish for México the gap is small enough that in check calls to drivers and screening interviews nobody has asked.
• Latency: self-hosted needs GPU capacity sized correctly. Undersized, the pauses get long. We size it during Blueprint and test it with real call volume before the gate.
• Operations: someone has to run the models, patch them and watch them. That is part of the retainer, not a surprise in month four.
• Model optionality: because the runtime is multi-model, a client can route the conversational step to Claude, GPT, Gemini or Nova when the reasoning demands it and keep the cheap path for the rest. The $0.035 figure is the fully open-weights path.

Where voice agents pay for themselves first

• Fleet check calls: a voice agent calls drivers twice a day, logs delays in the TMS and escalates exceptions. Typical target: delays detected in minutes instead of at the end of the shift.
• Screening interviews: about 1,000 calls an hour in our AI Interviewer case, each one transcribed and scored into the ATS (Proof Ledger, Sep 2026).
• Collections reminders with payment promises written to the core.
• Appointment confirmations and rescheduling connected to the PMS or the EHR.

How the price behaves at scale

Per-minute pricing is linear: ten times the calls, ten times the invoice. A self-hosted stack is a step function: ten times the calls, one more GPU node, and the per-call cost goes down. That is the property that lets a voice agent move from a pilot in one branch to the whole network without a new budget conversation. It is also the reason we insist on running the stack in the client's cloud: the capacity, the bill and the logs are yours.

How to compare quotes

Ask for the cost per 15-minute call with all three models included, at your volume, in your cloud. Ask which part of it is per-minute pricing you cannot negotiate. Ask what happens to the price at 10x. If the answer to the last question is a bigger invoice, the pilot will not scale. If it is a bigger GPU, it will.