What Does an AI Voice Agent Cost per Minute? The Full Breakdown

An AI voice agent minute has three cost lines. Self-hosted Dograh, the open-source platform we build, has no platform fee. Dograh Cloud charges a 1-cent-a-minute platform fee, Vapi 5 cents and Retell 5.5 cents. Model usage typically runs 5 to 12 cents. The phone line adds roughly 1 to 2 cents in the US.

Key Takeaways

• The phone line is a third cost line, and totals look low without it.
• Model usage moves the most, from about 3 cents to well past 12.
• Self-hosting removes the platform fee. Servers become a monthly line instead.

This post is part of our guide to The Business Case for AI Voice Agents: ROI and Payback, which names the three per-minute cost lines. Here we price each one with named rates and add them up, Dograh included.

The three lines on every minute

Every minute an AI voice agent spends on a call carries three separate charges, and a budget that counts only one or two of them comes out low.

The platform or orchestration fee pays for the software that runs the call flow, turn-taking and tool calls. Model usage covers speech-to-text, the language model and the voice, or one speech-to-speech model in their place. The phone line is the carrier's own per-minute rate, and it is the line most estimates leave out.

You pay all of them on every minute the agent is on the line. That includes the minutes before a call is transferred to a person and the minutes spent talking to a voicemail box, so both belong in your count.

Line one: the platform fee

The platform fee is the easiest line to compare, because each vendor publishes one rate. Self-hosted Dograh has no per-minute platform fee (servers, model keys and phone minutes still cost). Dograh Cloud has a 1¢-a-minute platform fee. Vapi charges $0.05 a minute for hosting. Retell charges $0.055 a minute for its voice infrastructure.

At 1,000 minutes a day, those rates come to $3,650 a year on Dograh Cloud, $18,250 on Vapi and $20,075 on Retell, before any model or phone cost. Self-hosted, the platform line is zero, and the cost moves to the servers you run.

Some products fold models into one bundled rate instead. Deepgram's Voice Agent API, for example, is priced at $0.075 a minute in its standard tier. A bundled rate is simpler to read and harder to compare, because you cannot swap one model inside it for another.

Line two: model usage

Model usage across speech-to-text, the language model and the voice typically runs 5 to 12 cents a minute, depending on the use case and the providers you pick.

Speech-to-text is usually the smallest part. Deepgram charges Nova-3 streaming at $0.0048 a minute for one language and $0.0058 for many, both marked as limited-time promotional rates. Its Aura-2 voice is $0.030 per 1,000 characters.

The language model and the voice swing the most. Retell's models run from $0.0016 for GPT 5 nano and $0.008 for GPT 5 mini up to $0.32 for GPT 6 Astra, and its voices run from $0.015 a minute to $0.10 for ElevenLabs v3 and v4. Vapi's estimates for its default Deepgram, OpenAI and ElevenLabs models add up to about 3.2 to 7.9 cents a minute together.

The line can sit below 5 cents when you run small or open models on your own GPUs. Cerebrium's own deployment ran speech-to-text, the language model and the voice for $0.02932 a minute, with no phone line counted. It climbs past 12 cents with a premium voice, a large language model or a speech-to-speech model.

Dograh is the orchestration layer that runs the call flow, tools and telephony, and the speech and language models are a separate layer you choose. In model configurations you pick the platform's own models, bring your own keys for a supported provider, or point it at open-weight models you host.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

Line three: the phone line

The phone line is the carrier's rate for carrying the call, and it is small per minute but never zero. Twilio charges $0.0140 a minute for outbound calls, $0.0085 for inbound calls to a local number and $0.0220 for inbound toll-free calls. A local number rents for $1.15 a month. Bringing your own SIP trunk to Twilio costs $0.0040 a minute instead.

Telnyx's call control pricing is $0.002 a minute plus SIP trunking from $0.0032 a minute, which puts it near half a cent. Its rates vary by destination and volume. Retell resells telephony at $0.015 a minute, and you can bring your own trunk to it at no charge.

Outside the US the numbers change, and calls to mobiles in some countries cost several times the US rate.

Dograh connects to your own carrier, such as Twilio, Telnyx, Plivo, Vonage or Asterisk ARI, so the phone line is billed by the carrier at its own rate.

Adding it up: one minute, line by line

Put the same providers under each platform, and only the platform line moves. The table prices one US outbound minute with the Deepgram, OpenAI and ElevenLabs models at Vapi's own estimates, and Twilio for the phone line. With your own keys you pay each provider's own rate, which can differ from those estimates.

Retell bills its models itself, so its minute is built differently. Voice infrastructure at $0.055, GPT 5 mini at $0.008, one of its $0.015 voices and its $0.015 telephony come to $0.093. Swap in an ElevenLabs v3 voice and the same minute is $0.178.

Some costs sit outside the minute. Phone numbers rent by the month. Each concurrent call line above the included number costs extra, such as $10 a month on Vapi and $8 a month for each one past 20 on Retell. Add-ons are billed per minute too, such as Retell's AI quality checks at $0.10 a minute after 100 free minutes. Self-hosting swaps the platform fee for servers you pay for every month. Our breakdown of Vapi pricing works through a full monthly bill.

Ways to bring the minute down

Most of the savings sit in the model line, so start there. Try a smaller language model and test it on real phone calls before you switch, because a model that passes in a demo can still fail on a hard call. Pair it with pre-recorded audio, which plays human clips where they fit and uses TTS only for the rest. Our post on cutting voice AI costs with pre-recorded audio shows how much that saves on the whole minute, which is far less than on the voice line alone.

Stop paying for minutes nobody hears. In Dograh, voicemail handling is off by default. Once you turn it on, "When voicemail is detected" is set to "Disconnect the call", so an outbound call that reaches voicemail ends there.

At steady volume, running open-weight models on your own servers can bring the model line down to GPU cost, and self-hosting removes the platform fee. That trade needs people to run the servers, which our post on Build vs Buy: Should You Build Your Own Voice AI Stack? weighs against renting a platform.