← Back to articles

Realtime voice economics

Voice agent API cost per minute: model, audio, and phone charges

A voice agent is not always billed at one price for each minute on the phone. Native speech-to-speech models may separately meter incoming audio, outgoing audio, text, and reused conversation context. A transcription → LLM → speech pipeline adds different services; a phone call may add carrier fees.

Updated November 11, 2026Provider-specific rates and units

Architecture first

Which services are on your bill?

ArchitecturePotential billable componentsCommon comparison error
Native realtime speech-to-speechAudio input, audio output, text/reasoning where applicable, retained conversation historyMultiplying call minutes by only the published fresh-audio estimate
STT → text LLM → TTSTranscribed audio, LLM input/output, synthesized speech, optional orchestrationCounting the LLM's token price as the whole stack
Phone-based agentEither speech architecture above, plus inbound/outbound call legs, numbers, and optional recordingAssuming the carrier's minute and the model's audio minute are the same billing unit

Browser-based voice agents may have no telephone carrier charges. A phone-based service can have both carrier and AI charges, and their billing boundaries need not align.

Worked example

Calculate audio usage, then account for missing charges

Google's published paid-tier example for Gemini 3.8 Live expresses fresh audio input as approximately $0.005 per minute and generated audio output as approximately $0.018 per minute. These are audio-token equivalents, not a universal flat session price.

2 min input audio

2 × $0.005 = $0.010

1 min output audio

1 × $0.018 = $0.018

Fresh audio subtotal

$0.028, not the entire session bill

This illustrative workload assumes those audio durations and the documented approximate conversion. It excludes text, thinking, repeated context, tool calls, platform fees, and phone charges. Google's Live API may bill retained audio conversation history again on later turns; its published per-minute equivalent should not be applied blindly to full call duration. Verify Gemini pricing and Live billing guidance.

Two different clocks

Phone-minute charges are separate

Twilio documents per-call-leg billing and rounding of partial voice minutes up to the next minute. Its published example of a 61-second inbound US toll-free call at $0.022 per minute costs $0.044 for the telephone leg alone. This is neither an AI-token price nor a universal telephony tariff; number type, country, direction, and product can change the carrier charge.

If a call is forwarded, more than one call leg may be billable. A browser microphone session without PSTN calling should not be assigned a hypothetical carrier price. Read Twilio's billing example.

Repeatable budget

Measure a representative call, not an assumed token-to-minute conversion

  1. Choose native speech-to-speech or a transcription + LLM + TTS pipeline. Identify each vendor and exact serving route.
  2. Run representative calls and capture the provider's actual input/output usage units, retained context, optional transcription, and any failed or repeated turns.
  3. Estimate model spend with the Voice / Realtime Calculator when its meter matches the model's billing unit; otherwise calculate separately using that provider's native unit.
  4. Add platform, carrier, recording, storage, and human-handoff charges only where the architecture uses them. Check applicable billing minimums or rounding rules.
  5. Track both total spend ÷ total billed conversation minutes and total spend ÷ successfully resolved calls. More minutes are not necessarily more value.

OpenAI's Realtime model card also distinguishes audio and text token meters. Do not substitute a text-input price for audio-input price. Published prices change; this guide is a method, not a quote for any individual deployment.