A8gent

Twilio ConversationRelay Pricing and Review (2026)

Twilio handles speech to text and text to speech inside the call and hands your server plain text over a WebSocket.

Reviewed by A8gent
Contents
  1. Verdict
  2. What is Twilio ConversationRelay?
  3. Best for, and not for
  4. Strengths and limits
  5. Pricing
  6. What people actually build with Twilio ConversationRelay
  7. What Twilio ConversationRelay actually costs at small-business volume
  8. The gotcha: $0.07 is the smallest part of what you have to buy
  9. What breaks in production, and how you find out
  10. What it connects to that you already run
  11. Twilio ConversationRelay alternatives: what to pick instead, and when
  12. Migration and lock-in: can you get your work out?
  13. FAQ
VerdictIs Twilio ConversationRelay worth it?

ConversationRelay is the cheapest way to add a voice agent to a Twilio number you already own, and a poor way to start from nothing. The rate is $0.07 a minute and the call is billed on top, so our reference workload of 4,000 local inbound minutes comes to $315.15 a month before the language model. What you get for that is Twilio doing the listening and the speaking while your server exchanges plain text. What you do not get is an agent: the model, the prompt, the tools and the server are all yours. Worth it for a team with a developer and an existing Twilio estate. Not worth it for a business that wanted a receptionist rather than a transport.

What is Twilio ConversationRelay?

Twilio ConversationRelay is a voice feature that turns a phone call into a text conversation for your code. Twilio transcribes the caller's speech, sends the text to a WebSocket server you run, and speaks whatever text your server sends back in a voice you choose. Your application supplies the intelligence: Twilio's product page says to bring your own language model. It is switched on with a single instruction in the call's TwiML, so it lives inside the same Twilio account, numbers and call flows a business may already have.

Best for

Teams already on Twilio that want a voice agent on an existing number without running an audio pipeline.

Not for

Anyone without a developer, or anyone who wants the language model, prompts and call flows managed for them.

Strengths

  • A single published rate of $0.07 a minute for the speech layer, on the same invoice as the call itself
  • Your server exchanges text rather than audio, so there is no audio bridge to build
  • Works on Twilio numbers and call flows you already have, including transfers to people
  • A choice of speech providers: Deepgram and Google for recognition, and Google, Amazon Polly and ElevenLabs voices
  • You choose and control the language model entirely, so an existing model contract carries over
  • Twilio publishes its own latency figures and their method, rather than an unqualified claim

Limitations

  • The language model is not included: you buy it, host the logic and pay for it separately
  • Call minutes, the phone number, recording and answering machine detection are separate lines on top of the $0.07
  • Twilio's price list gives the rate but does not itemise what it covers or say whether premium voices cost more
  • You must run an always-available WebSocket server, and validating Twilio's signature is your job
  • The AI and machine learning addendum must be accepted in the Console before it works
  • No builder, prompt manager or analytics: it is a transport, not an agent platform

Twilio ConversationRelay pricing

Twilio lists ConversationRelay at $0.07 a minute under intelligent service pricing on its US voice price list. That sits alongside, not instead of, the normal call charges: $0.0085 a minute to receive on a local number, $0.0140 to make a call, $0.0220 to receive on a toll-free number, and $1.15 or $2.15 a month for the number. There are no ConversationRelay plans or tiers; committed-use discounts are a sales conversation. Your language model is billed by whoever provides it.

Twilio ConversationRelay pricing tiers
PlanPriceWhat you get
ConversationRelay$0.07/minListed under intelligent service pricing on the US voice price list. Twilio provides speech recognition and text to speech; you bring the language model and a WebSocket server
Local number, inbound$0.0085/minPlus $1.15 a month per local number. Outbound calls from a local number are $0.0140 a minute
Toll-free number, inbound$0.0220/minPlus $2.15 a month per toll-free number. Outbound calls are $0.0140 a minute
Recording and detection$0.0025/minCall recording, plus $0.0005 a minute per month for storage. Answering machine detection is $0.0075 a call. Recording transcription is $0.0500 a minute
SIP, BYOC and in-app calling$0.0040/minSIP interface, bring-your-own-carrier trunking and browser or app calling are each $0.0040 a minute to make or receive
Committed useContact salesThe price list shows no volume tiers for ConversationRelay and says committed-use discounts are available by contacting sales

Pricing reflects public plans as of November 17, 2026 and can change. Check Twilio ConversationRelay for the latest.

What people actually build with Twilio ConversationRelay

ConversationRelay exists for one situation: a business or a software team already routes calls through Twilio and wants a language model to hold some of those conversations. It removes the audio engineering and leaves the agent to you. The builds that suit it follow from that.

The builds that fit

An after-hours answering agent on an existing number. The daytime call flow stays exactly as it is; outside opening hours the same number returns TwiML that connects the caller to ConversationRelay, and your server answers questions and takes a message. Nothing about the number, the carrier or the daytime routing changes.

A first-line agent in front of a human queue. The agent identifies the caller and the reason for calling, handles the simple cases, and hands the rest to a person with the context attached. Because this is ordinary Twilio call control, the transfer is the part that works most reliably.

Voice for a product that already has a text agent. If your support bot already runs on your own model and tools, ConversationRelay is the shortest path to letting it take phone calls: it sends your server text and speaks the text you return.

Outbound reminders and confirmations with a conversation attached. A local outbound call is $0.0140 a minute plus the $0.07, and answering machine detection at $0.0075 a call tells the agent whether to talk or leave a message.

Multilingual lines. Twilio's product page describes automatic language detection that responds in the caller's language, which suits a clinic or a letting agent serving a mixed community from one number.

What it is genuinely bad at

Being the whole solution. There is no prompt editor, no knowledge base, no tool catalogue, no conversation analytics and no hosted agent. A non-technical owner cannot configure it, and a developer still has to build, host and monitor a WebSocket service that is available every second the phone line is. If you are not already on Twilio, the appeal mostly disappears: you would be adopting a carrier, a programming model and a server in order to save a cent or two a minute against a platform that includes all three.

Bottom line: ConversationRelay is the right tool when the number, the call flow and a developer already exist and only the conversation is missing.

What Twilio ConversationRelay actually costs at small-business volume

Twilio's US voice price list, read on 17 November 2026, shows ConversationRelay at $0.07 a minute in its intelligent service pricing table. The same page lists the call itself separately: $0.0085 a minute to receive a call on a local number and $1.15 a month for the number. Twilio's own FAQ on that page says voice AI features carry additional charges, and we read the ConversationRelay line the same way: it is added to the call, not a replacement for it.

On our reference workload of 1,000 inbound calls averaging 4 minutes the arithmetic is short. ConversationRelay: 4,000 minutes times $0.07 is $280.00. The call leg: 4,000 times $0.0085 is $34.00. The number: $1.15. Total $315.15 a month.

The same volume on a toll-free number is more: 4,000 times $0.0220 is $88.00, plus $2.15 for the number, plus the $280.00, so $370.15. For outbound from a local number it is 4,000 times $0.0140, which is $56.00, plus $280.00 and $1.15, so $337.15, and answering machine detection on 1,000 calls adds $7.50.

If you record calls, add 4,000 times $0.0025, which is $10.00 a month, and storage at $0.0005 a minute per month, which is $2.00 for each month's recordings for every month you keep them. Keep a year of recordings at this volume and the storage line alone reaches $24 a month.

Then the line Twilio does not sell: the language model. Twilio's product page is explicit that you bring your own. What it costs depends entirely on the model and how much text each turn carries, so we are not putting a figure on it here; price it from your provider's token rates against a transcript of a real call. You also pay for wherever the WebSocket server runs.

Twilio ConversationRelay at 1,000 calls a month averaging 4 minutes (4,000 minutes). Rates from twilio.com/en-us/voice/pricing/us on 17 November 2026. Language model not included in any row.
Line itemRateLocal inboundToll-free inboundLocal outbound
ConversationRelay$0.07/min$280.00$280.00$280.00
Call minutes$0.0085, $0.0220 or $0.0140/min$34.00$88.00$56.00
Phone number$1.15 or $2.15/mo$1.15$2.15$1.15
Answering machine detection$0.0075/callNot neededNot needed$7.50
Recording (optional)$0.0025/min$10.00$10.00$10.00
Recording storage (optional)$0.0005/min per month kept$2.00 per month kept$2.00 per month kept$2.00 per month kept
Total without recording-$315.15$370.15$344.65
Language model and serverYour provider's ratesBilled elsewhereBilled elsewhereBilled elsewhere

What the $0.07 does and does not cover

This is the part the vendor's page does not answer. The price list gives a single rate and nothing else: it does not say on that row that speech recognition and text to speech are both inside it, and it does not say whether choosing an ElevenLabs voice rather than a Google or Amazon one changes the price. The documentation describes Twilio as providing both speech recognition and text to speech for ConversationRelay, and the price list shows no separate ConversationRelay speech line, so the reasonable reading is that the $0.07 is the speech layer in both directions. We could not confirm the premium-voice question from any public page on 17 November 2026. Before you standardise on a voice, run a day of test calls and read the usage report in the Console, which is the only authoritative answer.

Bottom line: At 4,000 local inbound minutes a month, ConversationRelay is $280 and the call is $35.15, so $315.15 before the language model you have to bring.

The gotcha: $0.07 is the smallest part of what you have to buy

The number on the price list is accurate and it is also the least informative number in the whole build. $0.07 a minute buys the listening and the speaking. It does not buy the thinking, and thinking is where both the cost and the work are.

Compare it honestly with a bundled rate. Deepgram's Voice Agent API is $0.075 a minute with a Standard-tier language model included. ConversationRelay at $0.07 looks half a cent cheaper and is not, because the model is still to come, and so is the server. Against a platform such as Bland AI that charges one rate with the model inside, the gap in the headline number is mostly the model you have not yet paid for.

The second trap is stacking. Because every Twilio feature is a separate per-minute or per-call line, a production configuration quietly accumulates them. The call leg, the number, recording, recording storage that grows every month you retain it, answering machine detection on outbound, and a transfer leg when the agent hands over to a person, which is a second call billed at its own rate. None is large. Together, on outbound with recording, they take the $280 to more than $350 before the model.

The third is time. ConversationRelay is charged per minute of the call it is attached to, so a caller who is put on hold inside the same session, or an agent that waits silently for a slow tool, is using billed minutes. A slow language model costs you twice: once in model latency the caller hears, and again in a longer call.

Finally, there are no published volume tiers for ConversationRelay. The price list says committed-use discounts are available by contacting sales and shows tiered pricing only for recording storage. A small business should plan on the list rate and treat any discount as something that arrives with real volume, not something to budget on.

Bottom line: ConversationRelay's $0.07 a minute covers the speech layer only, so it is not comparable with any rate that includes a language model until you have added your own.

What breaks in production, and how you find out

The component most likely to fail is the one Twilio does not run. ConversationRelay connects each call to a WebSocket server at a URL you supply. If that server is down, slow to accept connections or restarting during a deploy, the caller gets silence or a failed call, and Twilio has done its job correctly. That server needs the availability of a phone line, which is a higher bar than most small web services are held to. Run more than one instance, deploy without dropping open connections, and alert on connection failures rather than waiting for a customer to mention it.

Security is also yours. The onboarding documentation tells you to validate the X-Twilio-Signature header on the WebSocket handshake and accept only valid connections. Skip it and anyone who finds the URL can open sessions against your language model at your expense.

Latency is the next thing to watch, and here Twilio is unusually straightforward. Its product page states median latency under half a second and a 95th percentile under 0.725 seconds, footnoted as internal benchmarks (491 and 713 milliseconds) using different models, with results that may vary. Read the footnote properly: those are Twilio's figures under Twilio's conditions, and the largest variable, your language model and your server, is outside them. A large model with a long prompt will add far more delay than the speech layer does. Measure the gap between the caller finishing and the agent starting on your own stack, and keep the model small unless the conversation truly needs more.

There is also a setup failure worth knowing because it wastes an afternoon: ConversationRelay falls under Twilio's predictive and generative AI and machine learning features addendum, and it has to be accepted in the Console's voice privacy and security settings before the feature works.

Compliance, as Twilio words it

Twilio's product page FAQ says ConversationRelay supports HIPAA-eligible architectures and PCI-compliant workflows. That is carefully chosen language and should be read carefully: it describes what can be built, not a certification of what you built, and the page does not offer a business associate agreement on it. The model provider and your own server are both in the data path and both outside Twilio's statement. The documentation's notes on the speech providers also state that the customer is responsible for human review. Recording consent rules vary by state and country and remain your responsibility.

Bottom line: ConversationRelay fails when your WebSocket server does, so that server needs phone-line availability, signature validation and its own alerting.

What it connects to that you already run

The strongest integration is the obvious one: the rest of Twilio. A ConversationRelay session is started from TwiML, so it slots into existing numbers, call routing, queues and transfers without replacing any of them. Twilio's product page also mentions number provisioning and porting, so a number you hold elsewhere can be brought in rather than changed. For a business already running its phones through Twilio this is the entire argument for the product.

On speech, the documentation lists Deepgram and Google for recognition, and Google, Amazon Polly and ElevenLabs for voices, with ElevenLabs named at its Flash 2 and Flash 2.5 models. The product page names Deepgram Flux for speech recognition and turn detection and says you can use Twilio's built-in text to speech or bring your own.

On the model, anything. Your server receives text and returns text, so it can call any provider, a model you host, or an agent framework such as LangGraph or the OpenAI Agents SDK that you already use for chat. Twilio does not sit between you and the model at all.

For business systems there is nothing built in, and that is consistent with the design. Your server is ordinary code, so it calls your CRM, calendar and booking system directly or through function calling in whichever model you chose. The pattern we would recommend is to keep that server thin: let it hold the conversation and post a structured result to a webhook, and let an automation platform such as n8n or Make.com write to the systems of record, where the failures are visible and somebody other than the developer can fix them.

Bottom line: ConversationRelay integrates deeply with Twilio's own call control and not at all with your business systems, which your server reaches in ordinary code.

Twilio ConversationRelay alternatives: what to pick instead, and when

Sort the alternatives by how much they include. ConversationRelay includes the carrier and the speech layer. Everything else on this list includes either more of the agent or less of the carrier.

Pick the Deepgram Voice Agent API if you want the language model inside the per-minute rate. It is $0.075 a minute with a Standard-tier model and you still need a carrier, which in Deepgram's own guide is Twilio. Pick Cartesia Line if you want the agent code hosted for you as well: $0.03 a minute for the call, $0.014 a minute for telephony on a Cartesia number, model billed separately.

Pick Retell AI, Bland AI or Vapi if you want a platform with prompts, tools, analytics and numbers in one place and would rather not run a server. Pick ElevenLabs Agents if voice quality is the deciding factor, noting that ConversationRelay can itself use ElevenLabs voices. Pick LiveKit Agents if you want the full audio pipeline in your own code. The figures for those in the table come from our August 2026 reviews, not from today's check.

Alternatives at 4,000 minutes a month. Twilio, Deepgram and Cartesia rates read on 17 November 2026; other rows cite this directory's August 2026 reviews.
PlatformBilling unitAt 4,000 minutesWho it suitsDoes better than ConversationRelay
Twilio ConversationRelayPer minute plus call minutes; model is yours$315.15 local inbound, before the modelTeams already on TwilioCarrier and speech layer on one invoice
Deepgram Voice Agent APIPer minute of connection, model included$300 Standard, before telephonyDevelopers wanting one AI billLanguage model inside the rate
Cartesia LinePer minute, plus telephony and model$176 with Cartesia telephony, before the modelDevelopers who want hosted agent codeManaged runtime and a lower call rate
Retell AIItemised per-minute componentsAbout $354 on a lean configuration (August review)Teams wanting a builderPrompts, tools and numbers in one product
Bland AIOne bundled per-minute rateAbout $600 including the model (August review)Businesses wanting nothing left to priceModel included; no server to run
VapiPlatform fee per minute, providers at cost$0.05 a minute platform fee before providers (August review)Teams with their own provider keysA finished platform layer
LiveKit AgentsOpen-source framework; you pay providersDepends on the providers you assembleEngineers who want full controlEvery layer of the pipeline in your code

Bottom line: If you are not already a Twilio customer, a platform that includes the model and the hosting usually costs less in total than ConversationRelay plus everything it leaves out.

Migration and lock-in: can you get your work out?

The lock-in with ConversationRelay is to Twilio, not to the feature. The feature itself holds almost nothing of yours: the agent's logic, prompt, tools and conversation state all live on your server, in your code. If you switched the voice layer tomorrow, that server's reasoning would move unchanged and only the part that speaks Twilio's WebSocket message format would be rewritten.

What does bind you is the estate around it. Numbers, call flows written in TwiML, queues, recordings stored in Twilio and any SIP configuration are all Twilio-shaped. Numbers can be ported out, and that is the asset to protect, but TwiML is Twilio's own language and call flows written in it are rewritten rather than exported.

There is no contract in the way on list pricing: usage is pay as you go, so leaving is a matter of stopping. A committed-use agreement negotiated with sales changes that, and its term becomes the real lock-in. Do not sign one on a forecast.

The practical advice is the same as for every voice build, and ConversationRelay's design makes it easier to follow than most. Keep the conversation design in a document: what the agent says, what it must never say, how it confirms a name or a number, when it transfers. Keep the model and tools behind your own interface. Download recordings you need to retain to storage you control rather than paying per-minute storage indefinitely. Do those three things and the whole voice layer is a replaceable transport, which is exactly what Twilio is selling it as.

Bottom line: Your agent logic never leaves your own server, so moving off ConversationRelay means rewriting the message handling and re-porting the numbers, not rebuilding the agent.

Twilio ConversationRelay FAQ

How much does Twilio ConversationRelay cost?

$0.07 a minute, listed under intelligent service pricing on Twilio's US voice price list as of 17 November 2026. Ordinary call charges apply as well: $0.0085 a minute to receive on a local number, $0.0140 to make a call, and $1.15 a month for a local number.

Does the $0.07 include the language model?

No. Twilio's product page says to bring your own language model. Twilio handles speech recognition and text to speech, and your WebSocket server calls whichever model you choose and pays for it separately.

What does 4,000 minutes a month cost?

For local inbound calls: 4,000 times $0.07 is $280.00, 4,000 times $0.0085 is $34.00, and the number is $1.15, so $315.15 a month before the language model and the hosting for your server. On a toll-free number the same volume is $370.15.

Does ConversationRelay include speech to text and text to speech?

Twilio's documentation describes Twilio as providing both for ConversationRelay, with Deepgram and Google for recognition and Google, Amazon Polly and ElevenLabs voices. The price list itself does not itemise what the $0.07 covers or say whether a premium voice changes the rate, so check your Console usage report after test calls.

What do I have to build?

A WebSocket server reachable at a secure URL that receives the caller's words as text and returns the text to speak, plus the logic that calls your language model and tools. You also return TwiML that connects the call to ConversationRelay and validate Twilio's signature on the connection.

How fast is it?

Twilio's product page states median latency under 0.5 seconds and a 95th percentile under 0.725 seconds, from internal benchmarks using different models, and notes results may vary. Your own language model and server add to that and are usually the larger share.

Is ConversationRelay HIPAA compliant?

Twilio's wording is that it supports HIPAA-eligible architectures and PCI-compliant workflows. That describes what can be built with it rather than certifying your build, and your model provider and server are in the data path too. Confirm the contractual position with Twilio before handling health data.

Are there volume discounts?

The price list shows no volume tiers for ConversationRelay and says committed-use discounts are available by contacting sales. Plan on the list rate until you have volume to negotiate with.

Do I need to enable anything first?

Yes. ConversationRelay falls under Twilio's predictive and generative AI and machine learning features addendum, which must be accepted in the Console's voice settings before the feature will work.

Looking at alternatives? ConversationRelay gives you a text channel into a phone call and leaves the agent, the hosting and everything after the call to you. Autonoly works at the other end: it is a managed service for the workflow the call sets off, such as the booking, the record and the follow-up. If you have developers and a Twilio account, ConversationRelay is a sound transport. If you want the outcome without owning a WebSocket server, look at a managed route first. See the Autonoly review.