Research-backed calculator guide
AI Voice Agent Telephony Cost: Add PSTN, SIP, Numbers and Media to the AI Bill
Estimate AI voice agent telephony cost per minute by separating PSTN, SIP, phone numbers, media, recording, speech and model charges instead of trusting one bundled minute.
Quick answer
AI voice agent telephony cost per minute is not the same as the AI model’s cost per minute. A phone call can create a carrier charge for an inbound or outbound PSTN leg, a SIP or media-stream charge, a monthly phone-number fee, and optional recording, storage, transfer or analytics charges. The agent still has its own speech-to-text (STT), language-model and text-to-speech (TTS) meters. Add each meter in its native unit, then divide fixed monthly fees by connected minutes.
For a U.S. example dated November 2, 2026, Twilio’s published Voice table lists $0.0085 per minute for calls to local numbers and $0.0140 per minute to make U.S. calls. The same page lists $0.0044 per minute for Media Streams, $0.0025 per minute for recording, and $1.15 per month for a local number. These are vendor-published figures shown as of August 2026 on the page, not a universal market quote. [1] A direct-stack estimate must then add your chosen STT, LLM, TTS, storage and operational reserve.
The practical formula is:
All-in cost per connected minute = telephony + SIP/media + number allocation + recording/storage + STT + LLM + TTS + platform fees + reserve
A platform’s headline rate can already bundle some of those layers. For example, Retell displays separate voice infrastructure, LLM, TTS, telephony and add-on lines in its estimator. [4] Vapi similarly shows hosting, model providers and transport as separate parts of an estimate. [5] Always identify what “per minute” means before comparing it with a carrier rate.
What belongs in the telephony line?
1. PSTN direction and call legs
Start with the direction of the call. An inbound call has a leg from the caller through the carrier to your number. An outbound call has a leg from your number to the destination. A transfer can create a second leg: the original caller remains connected while the system calls a human or another destination. A forwarding design can therefore cost the inbound rate plus the outbound rate for the same conversational minute.
Twilio’s explanation is useful because it prices each leg and rounds partial minutes under its stated billing rule. Its worked example says a 61-second inbound call to a U.S. toll-free number is billed as two minutes at the listed $0.0220 per minute rate. [1] [2] This is why “call length” and “billable minutes” should be separate fields in a cost model. Other providers may bill per second or use different increments, so copy the rule from the provider actually on your account.
2. SIP trunking or media streaming
A SIP trunk carries signaling and voice between a carrier and an application or AI platform. A media-stream product can separately expose the audio to your agent over a WebSocket or another interface. Do not automatically add both a SIP rate and a media-stream rate if the vendor’s product bundles them; first map the architecture and invoice meters.
Twilio’s U.S. Elastic SIP Trunking page lists termination starting at $0.0011 per minute, local origination at $0.0034 per minute, and toll-free origination at $0.0130 per minute. It also lists numbers separately at $1.15 per month for local and $2.15 per month for toll-free. [2] These are different units from the Programmable Voice table. Use the rate card for the actual product path rather than mixing the lowest number from one product with the AI features of another.
3. Phone numbers and fixed monthly fees
A number is usually a fixed monthly cost, not a cost that disappears when calls are quiet. Convert it to a per-minute allocation only for comparison:
Number allocation per minute = monthly number fee ÷ connected minutes in the month
If one $1.15 local number supports 5,000 connected minutes, its allocation is $0.00023 per minute. At 50 connected minutes, it is $0.023 per minute. The same number therefore looks cheap or expensive depending on utilization. Model number count separately when each campaign, region or customer requires its own DID.
4. Recording, storage, transfer and analytics
Recording is an optional telephony-adjacent meter. Twilio lists recording at $0.0025 per minute and recording storage at $0.0005 per minute per month on the Voice page. [1] A transfer may be priced per invocation, per minute, or by the carrier leg. Analytics can also be charged per minute. Keep each as a separate row so a compliance decision does not get hidden inside an “AI minute.”
5. STT, LLM and TTS are not telephone charges
The telephone network transports audio. STT turns audio into text, the LLM generates decisions or responses, and TTS turns responses back into audio. Their billing units differ. Deepgram’s published page, for example, lists streaming STT per minute, TTS per 1,000 input characters, and its Voice Agent product per minute of WebSocket connection time. It also says normal audio requests are billed per second and that multichannel duration is summed across channels. [3]
That distinction matters for a voice bot. If TTS is billed per character, “one call minute” is not enough to calculate TTS. Estimate spoken characters or use the vendor’s measured usage export. If a managed voice-agent product bundles STT, a language model and TTS, treat its published minute as a bundle and do not add those same meters again.
Worked cost example
Assume a U.S. inbound support agent with these inputs:
- 1,000 connected calls per month.
- Five minutes per connected call, or 5,000 connected call minutes.
- One Twilio local number.
- Direct media streaming and recording enabled.
- Deepgram streaming Nova-3 monolingual STT at the published pay-as-you-go rate.
- Aura-1 TTS at 900 input characters per connected minute.
- An LLM planning assumption of $0.012 per connected minute. This is an estimate for the example, not a vendor quote.
- No transfer, storage, taxes, discounts, failed-call charges or support plan in the base case.
The table below is the visual demonstration for this guide. It shows the meter, the unit, the input and the calculation rather than presenting one unexplained blended number.
| Cost layer | Unit or assumption | Input | Calculation | Monthly output |
|---|---|---|---|---|
| Inbound local PSTN | $0.0085 / min | 5,000 min | 5,000 × 0.0085 | $42.50 |
| Media Streams | $0.0044 / min | 5,000 min | 5,000 × 0.0044 | $22.00 |
| Call recording | $0.0025 / min | 5,000 min | 5,000 × 0.0025 | $12.50 |
| Local phone number | $1.15 / month | 1 number | 1 × 1.15 | $1.15 |
| Deepgram streaming STT | $0.0077 / min | 5,000 min | 5,000 × 0.0077 | $38.50 |
| Aura-1 TTS | $0.015 / 1,000 chars | 4,500,000 chars | 4,500 × 0.015 | $67.50 |
| LLM planning estimate | $0.012 / min | 5,000 min | 5,000 × 0.012 | $60.00 |
| Base total | $244.15 | |||
| Base effective cost | 5,000 min | $244.15 ÷ 5,000 | $0.04883 / min |
The TTS row converts 5,000 minutes × 900 characters per minute into 4,500,000 characters, or 4,500 billing units of 1,000 characters. The 900-character assumption is workload-specific. Measure actual output because a concise receptionist and a verbose sales agent will not have the same TTS consumption.
For planning, add a 5% reserve for retries, short-call rounding differences, small usage variance and unmodeled operations. This reserve is not a provider fee. It adds $12.21 to this example, producing $256.36 per month, or about $0.05127 per connected minute. If you instead use Vapi’s no-success-package hosting price of $0.05 per minute, add that platform line only when Vapi is actually in the path; the illustrative total becomes about $0.10127 per minute before any Vapi package, taxes or other changes. Vapi describes model provider costs as pass-through and transport fees as provider-charged, so confirm the selected configuration before using this comparison. [5]
This example excludes a transfer. If 10% of calls transfer to a human for the remaining two minutes, calculate that second leg from its own outbound or termination rate. Do not multiply the original five-minute inbound average by a transfer price unless the provider bills that way. Record transfer count and transferred minutes as separate operational inputs.
How to compare bundled and direct-stack prices
A bundled platform is easier to purchase but harder to compare by headline rate. Retell’s estimator shows a sample cost with LLM, voice infrastructure, TTS and telephony lines, and its detailed table separately lists phone numbers and verified-number fees. Its FAQ also says the AI fee stops after a transfer while telephony continues for the transferred call. [4] That is exactly the kind of boundary a cost ledger should preserve.
A direct stack gives more control but exposes more meters. You may pay one carrier for numbers and PSTN, another service for streaming, a speech provider for STT and TTS, and an LLM provider for tokens. Deepgram’s page illustrates why unit normalization matters: STT is minute-based, TTS is character-based, and a managed Voice Agent option is connection-minute based. [3]
For a fair comparison, normalize each option to the same connected-call workload. Use the same call direction, average duration, percentage transferred, language mix, TTS character rate, recording policy and monthly number count. Then show fixed fees separately from variable cost. A provider quote should be labeled as a quote, while your LLM, character-rate and reserve assumptions should be labeled as estimates.
The model pricing comparison can help you keep model rates separate from the transport layer. For invoice validation after launch, use the AI API invoice reconciliation calculator to compare estimated meters with actual provider rows. If media, recording and model traces need operational attribution, the AI observability ROI calculator is a useful adjacent planning tool.
Run your scenario
Turn this framework into an editable estimate.
Use dated, source-linked assumptions and see cost per request, successful task and month.
Common exclusions and limitations
This worksheet is for estimating usage economics, not a quote or telecom contract. Rates can vary by country, destination type, carrier product, volume tier, committed-use agreement and account configuration. The U.S. figures above are dated to the pages read for this article; recheck the provider’s live rate card before committing budget.
Taxes, regulatory surcharges, emergency-calling charges, number porting, verification, implementation, support, concurrency reservations, compliance storage and professional services may be outside the per-minute rows. Failed calls and voicemail treatment also vary. Retell says it charges connected calls and tracks calls to the nearest second, while Twilio’s help example describes partial-minute rounding; those rules are provider-specific evidence, not a universal standard. [1] [4]
Finally, cost per connected minute is not cost per successful outcome. A short call that fails to qualify a lead can be less efficient than a longer call that completes a booking. Track connected minutes, transferred minutes, successful outcomes and human follow-up separately so a lower carrier rate does not hide a weaker workflow.
FAQ
Is PSTN the same as AI voice-agent cost per minute?
No. PSTN is the telephone-network leg. An AI voice agent can also incur SIP or media-stream, STT, LLM, TTS, recording, storage, number and platform charges. A bundled product may combine several of these, but the underlying meters remain conceptually distinct.
How do I calculate telephony cost for an inbound AI voice agent?
Multiply connected inbound minutes by the inbound rate for the number type, add any media or SIP rate, add optional recording and storage, then allocate monthly number fees across connected minutes. If calls transfer, add the second call leg using its own duration and destination rate.
What is the PSTN cost for a voice bot per minute?
There is no single universal PSTN rate. It depends on direction, country, destination or number type, carrier and product. In the U.S. Voice table read for this article, Twilio lists $0.0085 per minute for calls to local numbers and $0.0140 per minute for U.S. outbound calls, while its toll-free inbound rate is $0.0220 per minute. [1] Treat those as dated vendor-published examples, not a market average.
Is SIP trunk voice AI cost cheaper than a managed platform?
It can be, but the answer depends on utilization and the meters you must assemble. SIP may lower the transport line while leaving you responsible for numbers, media, STT, TTS, LLM, recording, monitoring and compliance. Compare the same workload and include fixed fees, concurrency and transfer behavior before concluding that one architecture is cheaper.
Should I add STT and TTS to a bundled voice-agent rate?
Only if the bundle excludes them. Deepgram’s Voice Agent rate is described as covering STT, a managed language model and TTS, while Retell and Vapi expose component lines in their estimators. [3] [4] Read the selected plan’s meter definitions and reconcile them to an invoice before adding another speech charge.
How should I model outbound calls and human transfers?
Keep outbound dials, connected outbound minutes and transferred minutes as separate fields. A failed dial may have different billing treatment from a connected call. A transfer may continue the telephony charge after the AI layer stops, so model the original leg and the human leg independently.
Primary sources checked
Pricing and product behavior can change. These sources support the time-sensitive claims in this guide; verify current provider terms before committing spend.