AI voice agent cost in 2026, from $49 a month to a $15,000 build

An AI voice agent answers your business phone, talks with callers, books appointments and hands the call to a person when it should. Working out the AI voice agent cost is harder than it looks, because the number a platform prints on its pricing page is a component price rather than a bill. Most articles on the subject quote a per-minute figure with no vendor attached and no date on it, which makes it worthless a few months later.

Every rate below was read off the vendor's own pricing page in August 2026 and attributed to that vendor. Where a page would not load for us, we say so and leave the number out.

Every figure below was read off a vendor page in August 2026

Vendor	Item	Published rate	Status
OpenAI	gpt-realtime-2.1 audio input	$32.00 per 1M tokens	Read live
OpenAI	gpt-realtime-2.1 cached audio input	$0.40 per 1M tokens	Read live
OpenAI	gpt-realtime-2.1 audio output	$64.00 per 1M tokens	Read live
OpenAI	gpt-realtime-2.1-mini audio in / out	$10.00 / $20.00 per 1M tokens	Read live
Twilio	Inbound local voice	$0.0085 per minute	Read live
Twilio	Outbound local voice	$0.0140 per minute	Read live
Twilio	US local phone number	$1.15 per month	Read live
Deepgram	Nova-3 monolingual streaming	$0.0048 per minute promotional, $0.0077 regular	Read live
ElevenLabs	Flash / Turbo low-latency TTS	$0.05 per 1,000 characters	Read live
LiveKit Cloud	Agent session	$0.0100 per minute	Read live
LiveKit Cloud	Inbound telephony	$0.0100 per minute	Read live
LiveKit Cloud	Bundled STT / TTS / LLM	$0.0058 / $0.0300 / $0.0014 per minute	Read live
Vapi	Platform per-minute rate	$0.05 per minute	Read live
Retell AI	Headline agent rate	$0.07 to $0.31 per minute	Read live
Twilio	A2P 10DLC brand and campaign fees	Figure unavailable	Help centre page would not load for us

One conversion underpins everything below, and it comes from OpenAI's realtime cost guide rather than a price list: "audio tokens in user messages are 1 token per 100 ms of audio" and "audio tokens in assistant messages are 1 token per 50ms of audio". That gives 600 input tokens per minute of caller speech and 1,200 output tokens per minute of agent speech, the bridge between a token price and a phone bill.

What a headline per-minute rate leaves out

Vapi puts $0.05 per minute at the top of its pricing page. Further down the same page it says model costs for speech-to-text, the language model and text-to-speech are charged separately at cost, or at zero if you bring your own key, and that telephony is "charged by provider, not Vapi". The $0.05 buys the hosting container. Everything a caller hears is billed elsewhere.

Retell AI is more inclusive and says so plainly. Its $0.07 to $0.31 per minute covers speech-to-text, the language model and text-to-speech, with the spread driven almost entirely by which model you pick: roughly $0.055 per minute of Retell voice infrastructure, $0.015 for TTS, and $0.003 to $0.16 for the model. Telephony sits outside it, and Retell's page cites $0.015 per minute for US calls through Twilio.

Neither vendor hides this; both disclose the exclusions on the same page as the headline. The trouble is that the headline is what gets quoted, and by the time the first invoice lands it has roughly doubled. The same pattern runs through automation tooling, which is why we wrote up the four meters behind a monthly automation bill.

How audio tokens turn into money

We run production voice on OpenAI Realtime with LiveKit for transport and Twilio for telephony. The model hears the caller and speaks back directly, with no transcription or synthesis step between, which is why interruptions and mid-sentence corrections feel natural. The cost is being billed in audio tokens, which behave differently from the per-minute meters people know.

Two things drive the bill. The first is simple: 600 tokens for every minute the caller speaks, 1,200 for every minute the agent speaks. The second catches people out. Every turn resends the accumulated history as input, so a call gets more expensive the longer it runs. OpenAI's guide puts it directly: "turns later in the session will be more expensive."

Caching stops that being ruinous, at $0.40 per 1M tokens against $32.00 uncached, an eighty-fold difference on the largest line in a long call. OpenAI calls caching "best-effort and not guaranteed" and warns that "removing or changing content in the conversation will 'bust' the cache up to the point of the change". If your agent rewrites its system prompt mid-call, or prunes old turns to save context, you pay in a way no per-minute estimate predicts.

Worked example: 300 calls a month at three minutes

A business takes 300 inbound calls a month averaging three minutes, so 900 connected minutes. On a three-minute call we assume the caller speaks for 45 seconds and the agent for 63, the remaining 72 seconds being silence, listening gaps, tool calls and the pauses either side of a transfer. That talk-time ratio of roughly 60 percent is worth checking against your own recordings, because an agent reading out four appointment slots talks far more than one taking a name and a number.

Component	Published rate, August 2026	Quantity per call	Cost per call
OpenAI Realtime audio input	$32.00 per 1M tokens	45s caller speech = 450 tokens	$0.0144
OpenAI Realtime audio output	$64.00 per 1M tokens	63s agent speech = 1,260 tokens	$0.0806
Replayed context, cached	$0.40 per 1M tokens	~10,800 tokens over 12 turns	$0.0043
Twilio inbound local	$0.0085 per minute	3 minutes	$0.0255
LiveKit agent session	$0.0100 per minute	3 minutes	$0.0300
Total per call			$0.1548

That is $0.0516 a minute all-in, $46.44 a month across 300 calls, and $47.59 with the $1.15 Twilio charges for the number. Swapping in the mini variant at $10.00 and $20.00 per 1M tokens takes the model lines from $0.0993 to $0.0322 per call and the month to about $27.90, a real option for simple message taking and a poor one for anything that reasons about a schedule.

Now break the cache. If those 10,800 replayed tokens were billed at the uncached audio input rate of $32.00 per 1M, that line goes from $0.0043 to $0.3456, the call goes to $0.4961, and the month goes to $148.83. Same traffic, same vendors, same published prices, three times the bill.

The cascaded pipeline bills on four separate meters

The alternative chains three models: streaming speech-to-text, a text language model, then text-to-speech. It is easier to debug, since you hold a transcript at every stage, and you can pick a specialist vendor per step. It also adds latency at every hop and handles interruptions worse, because the pipeline must notice the caller started speaking and then tear down a synthesis job already in flight.

Costed on the same call, with ElevenLabs Flash at roughly 900 characters per minute of speech:

Component	Published rate, August 2026	Quantity per call	Cost per call
Deepgram Nova-3 streaming STT	$0.0048 per minute promotional	Full 3 minutes of audio	$0.0144
ElevenLabs Flash / Turbo TTS	$0.05 per 1,000 characters	63s agent speech = ~945 characters	$0.0473
Language model	$0.0014 per minute, LiveKit bundled rate	3 minutes	$0.0042
Twilio inbound local	$0.0085 per minute	3 minutes	$0.0255
LiveKit agent session	$0.0100 per minute	3 minutes	$0.0300
Total per call			$0.1214

That is $0.0405 a minute, $36.42 a month, $37.57 with the number. Cheaper than speech-to-speech by about ten dollars at this volume, with two caveats. Deepgram flags that Nova-3 streaming rate as a limited-time promotion against a regular price of $0.0077 per minute, so the line could rise 60 percent with nobody doing anything wrong. And the STT meter runs on the whole call duration rather than on speech, because you stream continuously whether or not anyone is talking.

We picked speech-to-speech and paid the extra ten dollars. Callers interrupt receptionists constantly, and a pipeline needing 300 extra milliseconds to notice produces the stilted call that makes people hang up and dial a competitor.

Comparing the same 900 minutes across platforms

Setup	Platform fee	Model costs	Telephony and number	Monthly total
Direct on OpenAI Realtime, LiveKit, Twilio	$9.00	$29.79	$8.80	$47.59
Cascaded on Deepgram, ElevenLabs, LiveKit	$9.00	$19.77	$8.80	$37.57
Vapi at $0.05 per minute	$45.00	$29.79 at cost	$8.80	$83.59
Retell AI, cheap model end	$63.00	In platform fee	$14.65	$77.65
Retell AI, expensive model end	$279.00	In platform fee	$14.65	$293.65

The Retell rows use the $0.015 per minute US telephony figure from Retell's own page rather than Twilio's $0.0085, since that is what its documentation points customers at. The Vapi row assumes you bring your own OpenAI key, which its page says carries no markup.

None of this makes the platforms a bad purchase. Forty dollars a month buys a dashboard, call logs, a prompt editor and somebody else's on-call rotation. It stops being a good trade at volume, or the moment your call flow needs something the abstraction does not expose.

Meters the per-minute figure never shows you

• The phone number. Twilio publishes $1.15 a month for a US local number, which sounds like nothing until you run one per location for a fourteen-branch client plus spares for testing.
• Calls that never become conversations still burn telephony. Wrong numbers, robocalls, silent hangups at second three. Budget ten to fifteen percent above your answered-call count, then check the real ratio after a month.
• Somewhere to run the agent process. Self-hosting means a container or a VPS with a monthly floor whether or not the phone rings, and letting it scale to zero to save money buys you a cold start on a call that is already ringing, which a caller tolerates for about two seconds. We keep a warm replica, a fixed cost the per-minute maths never captures.
• Recordings and transcripts, plus the storage bill that grows monthly and the retention policy that decides when it stops. Consent law lives here too, so the decision is only partly financial.
• US outbound compliance: A2P 10DLC brand and campaign registration, plus STIR/SHAKEN attestation so your calls do not land tagged as spam. Twilio's help centre page listing those fees would not load when we checked in August 2026, so we quote no figure, and you should price it with your carrier first.

Then there is your own time. Every call the agent hands off is a call a human answers, and the handoff rate decides whether it saved you anything. An agent transferring forty percent of its calls is an expensive switchboard.

If your calls run longer

Duration is the biggest lever, so here is the same speech-to-speech stack at four average call lengths, holding 300 calls a month and the same talk-time ratio.

Average call	Minutes a month	Monthly total
2 minutes	600	$32.11
3 minutes	900	$47.59
5 minutes	1,500	$78.55
8 minutes	2,400	$124.99

Treat the longer rows as floors. Context replay grows faster than call duration, so one eight-minute call costs more than four two-minute calls even though the table prices them alike, and how much more depends on how well your cache holds. Measure it on your own traffic before signing a volume commitment.

Building it costs more than running it

Every number above is the running cost. The build is a separate line, and at these volumes it is larger by two orders of magnitude.

Our builds run $3,000 to $15,000 fixed, and care afterwards is $1,500 to $4,000 a month, both listed on the pricing page with no form to fill in. The spread is mostly integration count: one calendar plus message taking at the bottom, multi-location routing, CRM writes, payment collection and outbound callbacks toward the top. The money buys the work between a demo that impresses you and a system that survives a Tuesday, roughly the list on the production-ready checklist and the reason we quote fixed rather than hourly.

We run six systems in production, one of which is our own voice receptionist. It answers our line daily. In testing it double-booked a slot when two calls hit the calendar within the same second, and the fix was a lock on the booking write, one of the standard ways a voice agent drops calls and double-books and something no template lists as a feature.

Care is where the running story and the build story meet. Vendor prices move, models get deprecated, and a credential quietly expires at three in the morning, which has its own page here, being the most common way a working voice system stops working. What a retainer should and should not cover is set out in a separate post on care.

When a template subscription wins

A custom build is the wrong purchase for plenty of businesses, and the template tier is genuinely good now. Rosie publishes $49 a month for 250 minutes, $149 for 1,000 and $299 for 2,000, as read on its pricing page in August 2026. Goodcall publishes $79, $129 and $249 a month per agent, billed on unique callers rather than minutes, with $0.50 for each caller above the allowance.

Take the subscription when your calls are simple, meaning hours, directions, a message and one calendar. Take it when you get fewer than about a hundred calls a month, because a $3,000 build takes a long time to pay back at that volume. Take it when you do not yet know what your callers ask, since sixty days on a template tells you your real volume and where the script breaks.

Move up when callers start hanging up on the template, or when the calls it fumbles are worth more than the build.

Five questions for any voice vendor

• Which meters run during a call, at what published rate, and do they reach me at the provider's price or with a markup?
• Who owns the phone number, the OpenAI account and the telephony account if we part ways next year?
• How is the system tested before a deploy? A serious answer describes a fixed set of real call cases scored every time, rather than a demo call.
• Where does a call go when the agent cannot handle it, and what happens if that number rings out?
• What does a change to the call flow cost after launch, and how fast does it ship?

Any vendor should answer all five without checking with anyone. On our voice work the client owns every account, the meters run at cost, and bugs in the first thirty days are ours. The wider rate card sits in the 2026 agency pricing breakdown.

This post has a shelf life

Prices move. Deepgram's streaming rate is flagged as promotional on its own page today. OpenAI has cut realtime audio pricing more than once and will again. Any per-minute figure you read anywhere, this one included, is a snapshot with a date attached, and the date matters as much as the number.

These figures were verified on vendor pricing pages in August 2026. We re-check them quarterly and update this page rather than leaving stale numbers up, and if you spot one that has drifted before we do, tell us. To have the same arithmetic run against your real call log, the build sprint page explains how we scope voice projects.