TL;DR
- Twilio is the easiest recommendation for a team already operating its call-control stack. Choose Media Streams for your own audio stack or ConversationRelay when you want speech conversion handled in the relay.
- Telnyx has the lower public US transport components in this comparison, with explicit API, SIP and streaming charges. Its raw-audio interface fits a team that owns the voice application.
- Vonage is a strong existing-stack choice for NCCO call flows and WebSocket audio. Its configured country/account rates belong in the same workload formula; an unsupported fixed rate does not improve the buying decision.
- Compare the same destinations, directions and AI minutes. A cheaper carrier minute can lose its advantage if it requires more maintenance or an expensive migration.
Choose the layer first
There are two different voice purchases. Raw media gives your application audio to process and lets it send audio back. Your team chooses speech recognition, the model, speech generation and interruption handling. A speech relay converts caller speech into text and your application's response back into speech, leaving the application to own the conversational logic and tools.
A carrier API supplies neither your entire sales process nor a correct CRM mapping by default. It places the call, connects the media, reports events and supports the chosen transfer. A qualification result, booked appointment or resolved customer request is the application's outcome.
This comparison concentrates on voice architecture. SMS registration and message segmentation belong in the adjacent phone-and-SMS purchase; adding them to a voice-only budget would blur the decision. The worked scenario therefore uses five ordinary US local voice numbers, with no messaging costs.
The carrier comparison
US-dollar public bases checked on 30 September 2026:
| Dimension | Twilio | Telnyx | Vonage |
|---|---|---|---|
| Call control | Programmable Voice/TwiML | Voice API/Call Control, TeXML option | Voice API/NCCO |
| Own audio stack | Media Streams | Media Streaming with bidirectional audio | WebSocket endpoint through NCCO connect |
| US outbound base | $0.014/minute | $0.002 API + $0.005 example SIP = $0.007/minute | Destination/account rate |
| US local inbound base | $0.0085/minute | $0.002 API + $0.0032 example SIP = $0.0052/minute | Number/account rate; per-second voice billing |
| Ordinary local number | $1.15/month | From $1/month | Country/type/account rate |
| Raw-media charge used below | $0.0044/minute | $0.0035/minute | Selected configuration |
| Recording and storage used below | $0.0025/minute recording; initial storage $0.0005/minute/month | $0.002/minute recording; zero recording storage | Selected recording configuration |
| Primary fit | Existing Twilio application or speech-relay choice | New composed stack with a strong transport price case | Existing NCCO application and channel agreements |
Twilio's exact recording rate is $0.0025/minute, with initial recording storage at $0.0005/stored minute/month. Those are the rates used in the scenario. Telnyx's current voice example separates the API from SIP transport; its headline $0.002 is not the total call cost. Twilio US voice, Telnyx Voice API, Telnyx numbers, Vonage pricing basis.
Twilio
A natural fit for an existing TwiML application
Twilio Media Streams supports receiving call audio and, in bidirectional mode, sending audio back into the call. That is the relevant interface when an external voice stack owns speech and the model. The application must still coordinate the caller's interruptions, the model's output and any action it takes. Media Streams.
ConversationRelay is a different option. It handles speech recognition and generation while your WebSocket application exchanges text and controls the logic. This can remove some audio plumbing without surrendering the CRM rules or tool implementation. ConversationRelay reference.
For human handoff, a relay session can end with context and trigger a Connect action callback that routes the caller onward. Sending that context to the receiving rep is part of your application, not a reason to assume a transfer includes a complete CRM briefing automatically. Contextual handoff pattern.
Choose Twilio when your developers already know the call model or a useful application exists. It is less compelling when the new-build decision is driven primarily by US raw-transport cost and there is no implementation advantage. The ecosystem's practical value is reduced work, not a universal claim that its audio is faster or clearer.
Telnyx
A composed stack with lower listed US components
Telnyx's media interface can stream audio to your server and accept audio back. Its documentation supports bidirectional RTP and several codecs, including linear PCM for relevant AI integrations. A team with a model or speech-to-speech service can connect that stack to call control rather than buying a complete hosted agent. Media Streaming.
The economic attraction is the published composition. The API minute, SIP transport, raw media, recording and optional services are separately priced. This makes the lower US base meaningful, but also means a bill built from only the $0.002 API row is incomplete. Transfers and other invoked primitives add their own charges. Voice API rate card.
The current page also lists a Conversation Relay primitive separately. A buyer using that must price its selected configuration rather than apply the raw-media example below unchanged. Likewise, separate STT and TTS primitives are optional services, not mandatory additions to every external speech-to-speech design.
Choose Telnyx when your team wants the composed stack and lower public US transport components. It is a poor choice if its apparent saving comes from excluding SIP or media that the actual architecture requires. It is also a weak migration case at low volume if a functioning carrier application needs substantial rewriting.
Vonage
Strong fit for an existing NCCO workflow
Vonage's NCCO reference covers connecting calls, collecting input, recording and delivering speech or audio. A WebSocket connection can route live audio into a voice application. That supports the same broad engineer-owned architecture as the raw-media options above. NCCO reference.
The Numbers API supports selecting and configuring number inventory. Country and capability determine the usable number, while destination and direction determine the call price. Those units are more useful than a single worldwide “voice minute” figure. Numbers API.
Vonage's current public Voice API page describes pay-per-use and per-second billing. Its country/account rate presentation does not establish a complete fixed quote for this article's exact workload, so the scenario gives the quantities and cost formula. It does not adopt competitor-published Vonage rates as though they were the buyer's current contract. Vonage voice pricing.
Choose Vonage when the NCCO application, operating experience or wider communications agreement already fits. It is less persuasive for a price-only greenfield purchase requiring an immediately reproducible public US total. Keeping a suitable application can nevertheless be worth more than a small listed transport difference elsewhere.
One workload across the carriers
Assume five US local voice numbers and 3,000 completed calls/month:
- 2,000 outbound calls to ordinary US destinations, three AI minutes each: 6,000 outbound minutes.
- 1,000 inbound local calls, three AI minutes each: 3,000 inbound minutes.
- All 9,000 AI minutes use the selected raw media stream and recording, with one month of retained audio.
- Two hundred calls transfer to a human for another four minutes: 800 additional outbound-leg minutes.
This is an illustrative connected-minute workload, not a connect-rate forecast. The table first prices the original AI legs; the transfer line is added separately. Human-leg recording is not included.
| Original AI-leg component | Twilio raw-media design | Telnyx raw-media design |
|---|---|---|
| Five voice numbers | 5 × $1.15 = $5.75 | 5 × $1 = $5 |
| Outbound voice | 6,000 × $0.014 = $84 | 6,000 × $0.007 = $42 |
| Inbound voice | 3,000 × $0.0085 = $25.50 | 3,000 × $0.0052 = $15.60 |
| Raw media | 9,000 × $0.0044 = $39.60 | 9,000 × $0.0035 = $31.50 |
| Recording | 9,000 × $0.0025 = $22.50 | 9,000 × $0.002 = $18 |
| One month storage | 9,000 × $0.0005 = $4.50 | $0 published recording-storage charge |
| Communication component | $181.85 | $112.10 |
Add an assumed $0.06/AI minute for the external speech/model stack: $540. That makes the AI-leg subtotal $721.85 on Twilio or $652.10 on Telnyx. This model budget is an editorial assumption, not a carrier plan inclusion.
The human destination leg adds 800 × $0.014 = $11.20 to the Twilio example. For Telnyx, the example adds 800 × $0.007 + 200 × $0.10 transfer invocations = $25.60. The design must also account for any continuing original leg or other bridge/conference charges; those are outside these selected components. The resulting subtotals are $733.05 and $677.70 before such charges.
Now give either raw-media application eight maintenance hours/month at $75/hour, or $600, and allocate $2,400 setup over twelve months, or $200. The illustrated operating components become $1,533.05 for Twilio and $1,477.70 for Telnyx. These are a practical budget, not exact all-in invoices: taxes, further hosting, support, unsuccessful-call billing and bridge-specific charges can add costs.
Telnyx's difference is $55.35/month under these assumptions. A $3,000 carrier migration would take about 54.2 months to recover through that difference alone. Larger sustained volume improves the transport case; an existing reliable integration makes migration harder to justify.
For Vonage, apply its actual rate set to the same quantities:
five number rentals + 6,000 outbound minutes + 3,000 inbound minutes + 9,000 streamed/recorded minutes + selected storage + 800 human-leg minutes + transfer/bridge services + external model stack + maintenance/setup allocation.
Per-second billing can change the result for short calls, but this three-minute example does not manufacture a rounding advantage. The formula supports an existing-stack recommendation while keeping the destination and workload comparison honest.
Speech handling changes the comparison
Compare Twilio ConversationRelay against the raw-media design using the same 9,000 AI minutes. Its $0.07/minute charge is $630, with ordinary voice and number components of $115.25. Replace the external full speech/model budget with an assumed model-only $0.015/minute, or $135, because Relay already supplies speech conversion. Relay rate and separate voice basis.
Add the same $27 recording/storage and $11.20 selected human-leg component: $918.45. That is $185.40 more than the raw-media subtotal. If using the relay reduces maintenance from eight to four hours/month, the $300 capacity saving outweighs that increase. With the same $200 setup allocation, the illustrative relay design is $1,418.45, against raw media's $1,533.05.
Those maintenance figures are assumptions, not a claim that a relay guarantees half the engineering effort. They show the buying question: is the control gained from owning the audio stack worth the additional work? A team with an existing raw-audio stack may save no maintenance at all by moving to a relay.
Do not charge the relay and raw-media service for the same AI session unless the actual architecture uses both. Do not add a full external STT/model/TTS budget on top of a speech relay that already handles the selected speech conversion. Conversely, a raw media stream needs a real service to interpret and generate speech; the carrier rate is not enough.
Numbers transfers and finished tasks
Number ownership matters for return calls. Keep the number attached to the business workflow and give inbound calls a route when the original agent is unavailable. A five-number US inventory does not prove that the same number type is available in another country.
Country changes require a new rate basis. Calling a UK mobile from a US number is a UK-mobile destination purchase. It does not inherit the US outbound rate used here. Toll-free inbound, SIP and browser legs are also different units from ordinary local calls.
Audio suitability depends on the whole loop. Carrier media, resampling, speech service, model response time and playback buffering all affect the call. This article does not rank providers by an unmeasured latency number. A supported codec and a properly connected model are more useful buying evidence than a universal “fastest voice” claim.
A transfer is a state change, not just another dial. Preserve the caller's task ID, tell the receiving human what is needed, and record whether they actually accepted the handoff. An unavailable destination should lead to a callback or booking outcome, not a falsely completed task.
CRM completion is separate from call completion. Store the provider call ID and outcome against the business record. A transcript can explain the conversation, while the application records who owns the next action. This is especially valuable when an agent has answered correctly but the CRM write fails.
Which carrier to choose
Choose Twilio when the existing application or a suitable speech-relay design reduces implementation and operating work. Its price premium at moderate volume can be smaller than the value of familiar infrastructure.
Choose Telnyx for a new composed stack whose volume makes the lower listed US components valuable. Include SIP, media and transfer charges in the design, and preserve the control you actually need.
Choose Vonage for an existing NCCO workflow or communications agreement that fits the task. Price the same quantities using its selected rates rather than inventing a fixed public total.
Use the signal-based outbound playbook to connect the calling system to a useful next action. The infrastructure earns its place when the conversation reaches the correct outcome and leaves a record a human can act on.
FAQ
Is the lowest carrier minute the cheapest agent?
Not necessarily. Media, speech, the model and maintenance can exceed the carrier component. The same workload needs those services wherever it runs.
Are raw media and ConversationRelay the same purchase?
No. Raw media leaves the audio stack with your application. A speech relay performs speech conversion while the application controls the logic.
Why are human transfers separate in the example?
They add destination-leg minutes and can invoke transfer or bridge services. The AI's original conversation minutes do not describe the full human handoff.
Which is the price-led new-build choice here?
Telnyx has lower selected US components. That supports a new-build shortlist, but the difference alone may not justify rewriting an existing application.
Can Vonage still be recommended without a fixed public total?
Yes. Its documented call model and existing-stack fit are useful buying reasons. The quantity formula lets the configured rates be compared on the same job.
