Voice AI API costs are split by token type: GPT-Realtime-2 audio and text rates explained

Realtime voice models price audio and text tokens separately, and the audio rates dominate any voice product budget. OpenAI lists GPT-Realtime-2 at $32.00 per million input audio tokens and $64.00 per million output audio tokens, against $4.00 and $24.00 for text. A conversation that sounds cheap per call adds up quickly because every spoken exchange is billed at the audio rate on both sides.

The cached input rate is the one real lever for voice products: GPT-Realtime-2 lists $0.40 per million cached input tokens for both audio and text. A session with a large, fixed system prompt that is reused across turns moves most of the input cost to the cached rate instead of the full rate.

A practical budget method for a voice assistant: estimate average audio input seconds per session, audio output seconds, and the size of the reused system prompt; convert seconds to tokens at the provider's documented ratio or by sampling real sessions; then compute both the full-rate and cached-rate paths against the list prices.

Limitations: token-per-second conversion is not published as a fixed ratio, realtime sessions vary, and these are list prices collected on 2026-08-05. Measure real sessions before scaling, and re-check the pricing page before budgeting.