Realtime buys latency, not savings
Speech-to-speech models feel magical but audio tokens cost a premium. If cost matters more than latency, a cascaded pipeline wins — price it on the AI voice agent cost calculator.
What a minute of AI voice conversation costs
Speech-to-speech APIs bill audio tokens on both directions — roughly $0.06–$0.30 per conversation minute depending on model tier, which annualizes to painful numbers fast: a support bot handling 1,000 ten-minute calls monthly runs $600–$3,000 in model costs alone. The cascaded alternative (STT → LLM → TTS) costs $0.02–$0.08 per minute but adds 1–3 seconds of latency, which conversation UX research consistently flags as the threshold where users start talking over the bot. The realtime premium is a latency purchase — worth it for genuine dialogue, wasted on IVR-style menu flows.