Guide
What an AI voice agent costs per minute
An AI phone agent is a pipeline of four metered services running in a loop: speech-to-text (STT) transcribes the caller, an LLM decides what to say, text-to-speech (TTS) speaks the reply, and a telephony carrier connects the call. Each is billed differently (per audio minute, per token, per character and per call minute), which is why voice AI pricing is hard to compare at a glance. The calculator converts everything to a cost per minute. With the defaults above (500 calls a day, 4 minutes each), the stack costs about $0.0381 per minute, or $2,318 a month.
Speech-to-text (STT)
STT is billed per minute of audio, and the full call is transcribed whether or not the caller is speaking. Rates in this calculator range from $0.001 to $0.0077 per minute. Streaming matters more than price: batch transcription is cheaper, but a live agent needs partial transcripts within a few hundred milliseconds to feel responsive.
LLM
The LLM is usually the smallest line item because spoken replies are short, but it behaves like any multi-turn agent: every turn re-sends the system prompt and the conversation so far. At 4 turns a minute, a 4-minute call makes 16 LLM calls and re-reads 42,000 input tokens. Prompt caching keeps that cheap. Long calls with large system prompts and no caching are where LLM cost climbs. The LLM pricing calculator models the same context growth for text agents, and the LLM price table lists the per-token rates behind the LLM dropdown.
Text-to-speech (TTS)
TTS is billed per character generated, only for the share of the call the agent is talking. At about 825 characters per spoken minute and the agent talking 45% of the time, a call minute needs ~371 characters. That puts TTS between $0.00556 and $0.0371 per call minute depending on the voice, and premium voices are often the single largest cost in the stack.
Telephony and platform fees
Carriers such as Twilio and Telnyx charge $0.007–$0.014 per minute for US voice, and a WebRTC agent in a browser or app skips this entirely. Managed platforms such as Vapi and Retell add an orchestration fee of $0.05–$0.07 per minute on top of the underlying providers. That fee buys turn-taking, interruption handling, call transfer and analytics, which take real engineering effort to build. Use the platform field to compare a managed stack with a self-built one.
How to optimize voice agent latency and spend
Voice is the one LLM use case where latency matters as much as price. Callers expect a reply within about 800 ms of finishing a sentence, and the budget is shared across STT finalization, the LLM's time to first token and TTS time to first audio. The cheapest choices are often also the fastest:
- Pick a fast mid-tier or budget LLM. Flagship and reasoning models add seconds of thinking and cost several times more per token, with little benefit for scripted calls.
- Stream everything. Streaming STT, token streaming from the LLM and streaming TTS overlap the stages instead of running them back to back.
- Keep the system prompt short and cached. Fewer input tokens per turn means a faster first token and a lower LLM bill.
- Match the voice to the call. Use premium TTS voices for sales and brand-sensitive calls and cheaper voices for reminders, surveys and routine support.
- End calls cleanly. Detect voicemail and silence early, and hang up when the task is done; every idle minute still pays STT and telephony.
- Negotiate at volume. Past roughly 100K minutes a month, STT, TTS and platform providers all offer committed-use discounts.
At very high volume, self-hosting the LLM or a Whisper model on a dedicated GPU can be cheaper than per-token or per-minute pricing. The API vs self-hosting break-even calculator shows where that crossover sits, and the LLM VRAM calculator tells you which GPU you'd need.