Cost and self-hosting

Voice AI prices are quoted in many units: per character, per second, per token, per minute and per concurrent slot. These guides put every cost on the same footing, show the arithmetic and say when running your own models pays off and when it does not. Prices are as published in October 2026.

If you are buying minutes, AI calling agent cost per minute in India explains why published prices run from ₹0.80 to ₹25 a minute and adds up each line at your own call length. Arabic voice agent cost does the same in dirhams and riyals. Per-second STT billing shows why a silent stream still costs money, and cascaded vs speech-to-speech agents compares the two designs on language coverage, cost and control.

For self-hosting, read self-hosted voice AI cost first. Self-hosting usually costs more than APIs below about 1 million call-minutes a month and less above a few million; text to speech pays off first, speech to text second and the language model last or never. Critical utilisation gives the formula behind that order: a GPU must be busy about 5 to 6% of the time to beat an API for STT and TTS, but 48 to 84% for a language model. Base load and burst applies it to a hybrid fleet, and self-hosted Whisper vs an STT API redoes a popular cost argument for live calls.

For hardware, GPU prices for L4, L40S and H100 lists rental rates in the US and India, concurrent calls per 24 GB GPU explains what limits how many calls one card carries, and self-deploying open-source voice models sets out the order to move components and how to load-test before you commit.