How much does an AI voice agent cost per minute?

The five lines on every receipt

Every call is a sum of five things: streaming speech-to-text billed per audio minute, a language model billed per token, text-to-speech billed per character, the carrier's per-minute rate on phone calls, and whatever the platform charges to run it all. Here is the balanced stack, which is what most English phone agents should run.

Speech-to-text is $0.0065, the language model $0.0024 at about nine turns a minute with a 2,000-token prompt, and text-to-speech $0.0090 for an agent that speaks roughly half the call. Telephony is $0.012 on US numbers and $0 on web calls. The platform line is $0.010 on every minute.

What each default stack costs

Voiceflint ships six measured stacks. The totals below include telephony and the platform fee; the models column is what the providers charge at list price.

The lowest-cost stack (AssemblyAI Universal-Streaming, GPT-5 nano, Soniox TTS Real-Time v2) is $0.011 in models. The highest-quality stack is $0.055, a 5x difference for a call that most listeners will not tell apart on a phone line.

Where platforms add cost

The model lines above are the same everywhere; what differs is what sits on top. Orchestration platforms charge a per-minute fee, bundle models at a fixed rate, sell concurrency by the line, or all three.

Ten thousand minutes a month, same models

Take the balanced stack at $0.019 in models and run it for a month on each platform. Voiceflint adds $0.0029 per minute when it holds the provider keys, and nothing when you bring your own.

How to bring the number down

• Bring your own keys. The 15% managed-key fee disappears and the model lines show $0 on the receipt because the providers bill you directly. On the balanced stack that is $0.0029 per minute.
• Pick the model for the job, not the biggest one. GPT-5 mini handles booking, routing and FAQ calls; save the frontier model for calls where reasoning is the product.
• Cache the prompt. Providers with prompt caching bill cached input tokens at a fraction of list price, and a voice agent re-sends the same system prompt every turn.
• Use a flash text-to-speech model. Inworld TTS-2 Flash costs $0.0090 per minute against $0.029 for Cartesia Sonic 3.6, and answers faster.
• Prefer web calls where you can. The widget and browser calls skip the $0.012 carrier line entirely.

Questions

Are the 500 free minutes really free?

Yes. Every workspace gets 500 minutes a month on Voiceflint's keys with no card. After that you add credit and pay the per-minute lines shown on each receipt.

Do you bill silence and hold time?

Providers bill streaming speech-to-text on the audio minute, including silence, and Voiceflint passes that through unchanged. Text-to-speech and the language model only cost money when the agent speaks or thinks.

Does bringing my own keys cost more?

No. Your provider lines show $0 on the receipt and the 15% managed-key fee is dropped. You pay the providers at whatever rate you negotiated, plus the platform fee and telephony.