What does an AI voice agent actually cost?
Vendors quote a per-minute price. That number hides four separate bills. This works out the real one, using rates you can edit as the market moves.
Voice agent cost calculator
Every rate below is editable. Defaults are approximate public list prices — check your provider's current pricing before committing budget.
How to read the result
Three things surprise most people the first time they run these numbers.
| What people expect | What the maths says |
|---|---|
| The LLM is the expensive bit | It is usually the cheapest line on the bill. Inference prices collapsed; a four-minute call costs a fraction of a cent in tokens. Optimising your prompt to save tokens is nearly always wasted effort. |
| Text-to-speech is a rounding error | It is typically 50–70% of total cost. Realistic voices bill per character, and an agent speaking naturally emits roughly 700–900 characters a minute. This is the only line worth serious optimisation. |
| Building is obviously cheaper | On raw component cost, yes. Add engineering hours and the crossover usually sits near 10,000 minutes a month. Below that you are paying yourself poorly to avoid a platform fee. |
The constraint that actually kills voice projects
Cost is the question people ask first. Latency is the one that decides whether the thing ships.
You need first audio out within roughly 800 ms of the caller finishing their sentence. Past about a second, people start talking over the agent — a silent line reads as a dropped call, and the conversation falls apart. That budget splits three ways:
| Stage | Budget | Where it goes wrong |
|---|---|---|
| Speech-to-text finalisation | 100–250 ms | Endpointing that waits too long to decide the caller has stopped speaking. |
| LLM time-to-first-token | 150–400 ms | Large models and long system prompts. This is where fast inference providers earn their place over raw model quality. |
| TTS time-to-first-byte | 100–300 ms | Premium voices are often the slowest. The costly tradeoff in voice is frequently a latency one, not a price one. |
Note the tension: the voice quality that impresses in a demo is often the one that breaks the 800 ms budget in production. Pick for latency first, then see how good you can make it sound inside that ceiling.
Tools worth looking at
If you are building rather than buying, these are the pieces you will need. Links are region-aware where it matters.
Somewhere to put the thing
A voice agent needs a public endpoint for webhooks before it needs anything clever. Cheap shared hosting is fine to start — you are not serving the audio yourself.
Compare hostingWorkflow automation
What happens after the call — CRM writes, follow-up email, ticket creation. Worth separating from the call path so a slow CRM never adds latency to a live conversation.
See optionsRun your own numbers first
Before signing anything, put your real call volume into the calculator. Most voice projects are killed by a per-minute cost nobody modelled, not by the technology.
Back to calculatorQuestions
How much does an AI voice agent cost per minute?
A self-assembled stack typically lands between $0.03 and $0.09 per minute in raw provider costs. Managed platforms usually charge $0.10–$0.15, which bundles infrastructure, orchestration and support. Text-to-speech is almost always the largest single component.
Why is text-to-speech the most expensive part?
Speech-to-text and LLM inference have both fallen sharply in price. Premium neural TTS has not, because it is still the hardest part to run cheaply at low latency. It bills per character, and natural speech runs 700–900 characters a minute, so it commonly accounts for 50–70% of total cost.
Should I build my own stack or buy a managed platform?
Component costs are lower if you build, but that ignores engineering time, latency tuning, call orchestration and being on call when it breaks at 2am. Below roughly 10,000 minutes a month, a managed platform is usually cheaper once your own time is priced honestly. Above that, building starts to pay back. Set the engineering fields in the calculator to your real numbers and it will tell you which side of the line you are on.
What latency do I need for it to feel natural?
Under about 800 ms from the end of the caller's speech to first audio out. Past a second, callers talk over the agent. This is a harder constraint than cost and should drive your model and voice selection before price does.
Do these numbers work for India and other non-US regions?
The STT, LLM and TTS rates are global — those providers bill the same regardless of where you are. Telephony is the exception and varies substantially by country and carrier. Replace the telephony rate with your actual local per-minute cost and the rest of the model holds.
Are outbound calls priced the same as inbound?
The AI components cost the same. Telephony differs, and more importantly outbound carries regulatory weight that inbound does not — consent and calling-hours rules under TCPA in the US, Ofcom rules in the UK, and equivalents elsewhere. The compliance cost of outbound is usually larger than the per-minute cost. Model that before you model this.