How Much Does Retell AI Cost in 2026? Pricing, Problems and Fixes

Retell AI is a developer platform for building AI phone agents, billed by the minute: $0.07 to $0.31 a minute blended on pay-as-you-go, with $10 in free credits and 20 free concurrent calls, though a real call starts around $0.087 a minute once an LLM and telephony are added (retellai.com/pricing, checked 24 September 2026). It is infrastructure you build on rather than a finished contact-center product, and it expects an engineer to own the pipeline.

What is Retell AI?

Retell AI is a voice AI platform that lets companies build agents to handle inbound and outbound phone calls: receptionist work, appointment booking, lead qualification, outbound campaigns, and call routing to a human when the conversation needs one. It came out of Y Combinator, was co-founded by Evie Wang, and by TechCrunch's 2024 profile, had raised a reported $4.53 million at the time. Retell announced a $4.6 million seed round led by Alt Capital, and Sacra puts total funding at about $5.1 million. That is a small amount of capital next to the revenue Retell is now reported to generate: Sacra estimates about $80 million in annualized revenue as of August 2026, up from roughly $45 million at the end of 2025.

The target buyer is explicit in how the product is built: engineering teams and technical founders who want to assemble a voice agent from parts, a speech-to-text vendor, an LLM, a text-to-speech voice, a telephony provider, instead of buying a packaged call-center seat. Industries named on Retell's site include healthcare, financial services, insurance, logistics, retail, and travel, and the homepage displays customer logos including Anker, Motorola, Grab, Lenovo, and Capsule alongside smaller healthcare and services brands.

Three case studies on Retell's site give a concrete sense of what a working deployment looks like: Pine Park Health, a senior-care provider, reports a 38% increase in scheduling NPS after deploying Retell for appointment coordination. SWTCH, an EV-charging company, reports cutting support costs by more than half. Medical Data Systems reports its Retell agent handling all of its inbound calls, transferring roughly 30% to a human, and collecting about $280,000 a month through the automated flow. These are vendor-published figures we could not independently audit, but they are specific enough to be checkable if you ask Retell for the underlying account during a sales call.

How does Retell's voice pipeline work?

Retell is built around a voice pipeline it assembles from named, swappable components rather than one black-box model:

• Speech-to-text: turns the caller's audio into text the LLM can reason over, tuned for the lower audio quality of a phone call instead of a clean studio recording.
• An LLM of your choice: Retell's pricing page lists model options from a "GPT 5 nano" tier at $0.0016 a minute up to "GPT 6 Astra" at $0.32 a minute, or $0.64 on its fast tier. The model is a cost lever you set per agent.
• Text-to-speech: Retell's own platform voices, priced separately from third-party voice providers.
• Turn-taking and latency management: the part Retell invests in most heavily by its own account. The company markets roughly 600ms round-trip latency and a proprietary turn-taking model, aimed at stopping the agent talking over a caller or leaving the awkward pause that makes a call feel robotic.

Telephony connects through SIP trunking, Twilio, Vonage, or a number Retell sells directly, which is what lets an existing business phone number route into an AI agent instead of a human queue. Agents are built through a visual, no-code flow builder for the conversation logic, but every layer underneath, model, voice, transcription vendor, is a configuration choice, which is closer to assembling infrastructure than filling out a chatbot builder's forms.

That component-level control is also where the pricing complexity comes from, which we get to below. Retell's own build documentation walks through choosing an agent type, then configuring speech recognition, LLM behavior, voice, and telephony behavior as four separate decisions rather than one setup screen.

What features does Retell AI include?

• Voice and chat agents from one platform: the same builder produces phone agents and web/embedded chat agents, so a team building one conversation flow can reuse it across channels.
• Simulation and testing tools, including an LLM playground and simulation testing before an agent takes a live call.
• Live call monitoring and a custom analytics dashboard that charts success rate, latency, cost, and concurrency, with filters and breakdowns you build yourself rather than a fixed report.
• AI quality assurance that can analyze up to 100% of calls instead of a sampled subset, billed separately at $0.10 a minute with the first 100 minutes free.
• Compliance certifications: HIPAA, SOC 2 Type II, GDPR, and ISO 27001, plus an enterprise option for on-premises deployment, which is why healthcare and insurance show up repeatedly among its named customers.
• Batch and scheduled outbound calling, for lead qualification and outreach campaigns on top of inbound support.

How much does Retell AI cost?

Retell publishes real numbers, which is rarer than it should be in this category, but the total cost is assembled from several line items rather than one plan price. We checked retellai.com/pricing directly on 2026-09-17 and again on 2026-09-24; the numbers below had not changed.

Pay-as-you-go starts at $0, with $10 in free credits and 20 free concurrent calls to test with. Beyond that, a voice call costs a blended $0.07 to $0.31 a minute, built from:

On top of the per-minute meter, phone numbers cost $2/month, a verified number costs $10/month, a US toll-free number costs $5/month on Twilio (Telnyx sells US local numbers only) with toll-free inbound at $0.06/minute, and each extra unit of call concurrency beyond the free 20 costs $8/month. Conductor, Retell's build copilot, is free to 30 messages per user and 200 per workspace a day, then $0.20 a message if an admin turns on pay-as-you-go and paused if they don't. Enterprise pricing is custom and adds dedicated implementation support, custom MSA/DPA terms, role-based access control, custom SSO, and a 24/7 support portal.

Add the components up before you take the headline range at face value. The $0.07 floor is exactly $0.055 of voice infrastructure plus $0.015 of platform text-to-speech, and nothing else. It assumes no LLM and no telephony, which is to say it assumes a call that neither thinks nor connects to a phone. The cheapest configuration that actually places a call is $0.055 + $0.015 + $0.0016 + $0.015, or about $0.087 a minute, and that's on the cheapest model with the cheapest voice. Swap in an ElevenLabs voice and you're at $0.112 before the LLM gets interesting. The range isn't dishonest, but the bottom of it isn't a price anyone pays.

The incentive worth naming here: because the LLM line item alone can be 1 to 400 times more expensive depending on the model you pick, a team can genuinely halve its per-minute cost just by choosing a cheaper model tier, which most packaged contact-center platforms never expose as a lever at all. That transparency rewards a team that tunes its stack and quietly punishes one that leaves the default model running at scale without checking the bill.

At a support-relevant volume, say 4,000 minutes of calls a month (roughly 800 five-minute calls), a mid-tier model landing around $0.15/minute all-in works out to about $600/month before phone numbers and add-ons. That is a useful anchor if a sales call quotes you something wildly different.

What do real users say about Retell AI?

Retell's G2 reviews page returned a 403 to an automated fetch on 2026-09-17. Read in a browser on 2026-09-18 it showed 4.8 out of 5 from 2,638 reviews, the largest verified review base of any voice platform we've priced. For the reviews themselves we used Trustpilot. Retell holds a 4.9 out of 5 rating from 814 reviews there (unchanged on 2026-09-24), 802 of them posted in the preceding twelve months. The distribution is 94% five-star, 5% four-star, and 1% one-star, with almost nothing in between. Trustpilot skews toward people motivated enough to leave a review either way, and a shape like that, a wall of fives with a thin spike at one, is what a self-serve product with a fast onboarding and a thin support tier usually looks like.

The praise clusters around two things: ease of building an agent and the pace of updates. One reviewer, Anatoli, wrote: "I like that they update the platform regularly and each update actually makes things easier and more flexible to use." Danny M's review is shorter and points at the same thing: "Retellai is the best. literally just type what you want the agent to do and the ai build the agent for you." Christian Landsteiner (February 2026) points somewhere more operationally useful: "Where Retell really stands out compared to other providers is reliability and uptime."

The one-star spike is not random, and it has a single theme: support after you've paid. M Saeed (July 2026) writes "I have very bad experience with this AI. no Support no one here to help after making payment." Abdulselam Almnini (March 2026) is more specific about the same gap: "No clear support channel, no timeline for resolution, no transparency about which countries are restricted and why." That's roughly eight one-star reviews out of 814, so it isn't a widespread failure, but two independent reviewers four months apart describing the same missing escalation path is a pattern rather than an outlier. It's also the predictable shape of developer infrastructure sold self-serve: the product is fast to start and there is no account manager to call when a call won't connect in a country you needed.

What goes wrong with Retell AI, and how do you fix it?

Retell runs a public support forum at community.retellai.com, which is the most useful thing about evaluating it: you can read what breaks and what the fix was before you sign up. We read it alongside the one-, two- and three-star Trustpilot reviews on September 19 and 20, 2026. Six problems dominate, and four of the six have a documented fix.

1. The per-minute bill is higher than the calculator said

The line item that surprises people is not on the pricing page. Retell's billing-exceptions doc says a prompt over 4,000 LLM tokens scales the billed duration of the whole call, not just the model line: "Scaling Factor = Prompt LLM Tokens ÷ 4,000" and "Billed Duration = Original Duration × Scaling Factor (rounded up)." A 5-minute call on an 8,000-token prompt bills as 10 minutes of everything, telephony and voice included.

What counts toward those 4,000 tokens is broader than the system prompt: the global prompt, function descriptions, state and node prompts, the call transcript so far, tool history and knowledge-base content. Flex mode compiles all of it into one context, which is why it triggers the surcharge most often. The threshold has moved before, from 3,500 tokens when token-based LLM billing launched on May 16, 2025 to 4,000 in the current docs, and the forum thread Token Surcharge Calculation (May 20, 2026) works through the arithmetic. A second thread, Unexplained 30% token usage increase (March 15, 2026), reports average consumption rising from about 2,800 to about 3,600 tokens a call with no configuration change, which is how a stable agent drifts across the line.

There is a second exception worth knowing: a call shorter than 10 seconds that uses a dynamic opening message carries a minimum charge of 10 seconds. On a high-volume outbound campaign with a lot of immediate hang-ups, that adds up.

Fix: check the token count on the LLM line of your call log, not the prompt you wrote, because the transcript and tool history grow inside it. Trim the global prompt to what applies to every turn, scope the knowledge base so it returns less, and move branching logic out of one long prompt. Retell's prompt-engineering guide recommends sectional prompts and switching to Conversation Flow when you are "coordinating many tools, managing state across conversations, or experiencing inconsistent single-prompt behavior," which also keeps the per-turn context smaller.

2. Transfers fail, and extensions fail more

Call transfer is the single most common topic on Retell's forum, with threads running from Transfer call function doesn't work (December 30, 2025) through Call transfer failures (January 25, 2026) to Transfer call doesn't work (April 8, 2026). The usual error, quoted in Issue with call transfer (January 12, 2026), is "Transfer call cannot be completed, telephony provider or carrier declined or was unavailable," which points at the carrier rather than at your agent.

The fixes Retell's own support gives, in the order they appear: switch the transfer mode from SIP REFER to SIP INVITE and confirm with your SIP trunk provider that REFER is supported at all; switch from cold to warm transfer; use the agent's number as the caller ID instead of the caller's; or add custom SIP headers.

Extensions are their own case. It doesn't transfer my calls to phone extensions (January 29, 2026) and Call Transfer to Extension not working when Displayed Caller ID is set to User's Number (August 3, 2026) both trace to the same two settings: a Displayed Caller ID of "User's Number" stops the extension working, and the DTMF digits get eaten by answering-machine detection, so support's fix is to set optOutHumanDetection: true on the transfer_call tool or use a cold transfer. Extension digits must be 0-9, * or #, written as 123#.

Fix: test a transfer to a real extension on your own trunk before launch, with Displayed Caller ID set to the agent's number and human detection off on the transfer tool. If REFER is not supported by your carrier, use INVITE and plan for warm transfer's overhead, covered next.

3. Latency is two to four times what the agent is configured for

Retell markets roughly 600 ms round-trip, and the forum is where the gap between that and production shows up. Slow Latency (3000ms, triple what it should be) (January 23, 2026) reports "around 3000ms latency despite my agent being configured for 770-1100ms." Latency issues inbound calls (January 12, 2026) opens on end-to-end latency above 2,500 ms. Rigid Mode Latency Issue (January 20, 2026) measures about 5,000 ms in Rigid mode against about 1,300 ms in Flex on the same agent. And Warm transfer additional latency (May 1, 2026) records "an additional 1-2 seconds of latency between the end user and the transferred to agent" after Retell's agent drops off. On Trustpilot, David Hartmann (UAE, March 25, 2026) put the consequence plainly: "the latency of the model didnt support fast paced conversations."

The pipeline is assembled from parts, so the delay has a specific owner on every call: speech-to-text, the model, text-to-speech, telephony, and the agent mode you chose. The mode is the one people miss, because Rigid and Flex are a design decision that turns out to be a latency decision.

Fix: read the estimated-latency figure in the builder against the measured LLM response time in the call log, and treat the difference as your model's problem and not the platform's. Move to a lighter model tier for the turns that do not need reasoning, keep the prompt short (which also keeps you under the surcharge in problem 1), and if a queue needs fast turn-taking, test Flex before committing to Rigid. Budget the extra second or two that warm transfer costs into your handoff script, so the customer isn't left sitting in silence.

4. Outbound calls are blocked until verification clears, and that takes weeks

Buying a number does not enable outbound calling. Outbound calls blocked, "No valid KYC" even with number and live agent (January 28, 2026) gets the direct answer: "Purchasing a phone number alone doesn't automatically complete KYC verification. You need to complete the verification process separately." Pending Processing on Verified Phone Number (January 16, 2026) puts the timeline at "1–2 weeks after you submit your business profile details," and the poster was 24 days in. Outbound Calls Blocked, Persona Verification Failed (January 14, 2026) covers the identity check failing repeatedly, and Unable to verify phone numbers at sign up for a Nigerian number (January 13, 2026) covers not getting past signup at all.

The Trustpilot version is angrier about the same process. Karen Róbertsdóttir (Iceland, August 5, 2025) gave up when, after buying a number, "they wanted my f'ing passport." Shefeeque Abdul Rahman (Malaysia, March 7, 2025) wrote that Retell "will block your account without telling you ahead of time," for a week.

Fix: submit the business profile and the A2P 10DLC registration in week one of the build, not in the week you planned to launch, and assume one to two weeks. Identity verification is a US telecom requirement rather than a Retell policy, which is worth knowing before you read it as obstruction.

5. Calls to some countries simply do not connect

Abdulselam Almnini (Turkey, March 11, 2026) quoted the error he got dialing Turkish numbers: the destination "is not included in whitelisted countries USA, CAN." He bought his own Telnyx number to get around it. Homecracker (UK, July 29, 2025) called the platform "unusable in the UK."

The mechanism is in Retell's international calling docs: a destination "just needs to appear in the supported list," and that list depends on the carrier behind the number you bought and not on your plan. Retell's own telephony pricing is inconsistent enough to check twice: the international docs list Telnyx US at $0.03 a minute against Twilio US at $0.015, while the pricing calculator labels a single combined option "Retell Twilio/Telnyx" at $0.015.

Fix: check the supported list for your carrier against your actual destination countries before you build, and bring your own Twilio or Telnyx number where Retell's default carrier does not cover them. US toll-free numbers are $5 a month and are Twilio-only, with toll-free inbound at $0.06 a minute; Telnyx sells US local numbers at $2.

6. Support after payment, and what has changed since those reviews

The one-star tail on Trustpilot is almost entirely one complaint. M Saeed (July 2026): "I have very bad experience with this AI. no Support no one here to help after making payment." Leshawn Clarke (UK, December 22, 2025): "they have an AI that replies to you instead, and an AI cannot help with verification. Also they tell you to get support on discord." Basti (Germany, October 8, 2025): "you need to join a Discord as their help desk and they never reply." Abdulselam Almnini again (March 2026): "No clear support channel, no timeline for resolution, no transparency about which countries are restricted and why."

That is eight or so reviews out of 814, and the picture has moved since most of them were written. Retell's support doc now names a support portal and [email protected], and says the team is "on call 7 days a week, with active hours 9 AM to 9 PM PST." The community forum every link above points at is public and searchable, and there is a status page at status.retellai.com.

Fix: search the forum first, because most of the problems on this page already have a resolved thread. Use the portal rather than Discord for anything account-related, and subscribe to the status page so an outage is not something you diagnose in your own agent for an hour. If your use case cannot tolerate community-first support with no named contact, the Enterprise tier is where the dedicated implementation support and the 24/7 portal live, and that is a sales conversation rather than a plan you can buy online.

What are Retell AI's pros and cons?

Pros

• Published, itemized pricing you can model yourself instead of a sales-gated quote.
• Real component-level control over the model, voice, and transcription vendor, which lets a technical team tune cost and quality independently.
• Genuine omnichannel reuse: the same agent logic runs for phone and embedded chat.
• Compliance certifications (HIPAA, SOC 2 Type II, GDPR, ISO 27001) that support regulated use cases like healthcare scheduling.
• A 4.9/5 Trustpilot score across 814 reviews is a meaningfully large, positive sample.

Cons

• The all-in per-minute cost is assembled from up to eight separate line items, which makes a clean monthly estimate harder than a flat-rate competitor. The advertised $0.07 floor excludes the LLM and the telephony, so the realistic cheapest call is about $0.087 a minute.
• It's a builder's tool. A team without engineering capacity to configure the pipeline will likely need Retell's paid implementation support or a systems integrator.
• Support after the sale is the single recurring complaint in an otherwise strong review base, and it's the whole of the one-star tail rather than a scattering of unrelated grievances.
• Retell handles voice calls. It has no ticket queue, no email channel, and nothing that resembles a help desk, so it solves one channel of support, not the whole inbox.

Who is Retell AI best for?

Retell fits a team that already has, or is building, phone as a real support or sales channel and has engineering resources to own a voice pipeline: choosing a model, tuning latency, wiring telephony, and watching a per-minute bill. Healthcare scheduling, home-services dispatch, and outbound lead qualification are the shapes of business the case studies describe, and the compliance certifications back up a real appetite for regulated industries.

It's the wrong choice if your support volume mostly lives in email and chat tickets rather than phone calls, or if you want a packaged product with a single monthly price instead of a metered infrastructure bill you have to model yourself. A 15-person e-commerce support team fielding Zendesk tickets has almost no use for a per-minute voice pipeline; a 15-person home-services dispatcher fielding 2,000 inbound calls a month has a very direct one.

How does Retell AI compare with Vapi and Macha?

The closest comparison to Retell is Vapi, another developer-first voice AI platform with the same assemble-your-own-pipeline model (vapi.ai/pricing, checked 24 September 2026). Vapi's free tier gives $5 of credits and $0.05/minute hosting; its Core plan is a flat $29/month for small teams; Pro starts at a $999/month minimum plus 10% of hosting fees; Premier is custom. Vapi's own calculator estimates $82–$129/month for 1,000 minutes at typical transcription, model, and voice settings, a genuinely different cost shape from Retell's blended $0.07–$0.31/minute. Run both calculators against your real call volume; a headline rate from either vendor won't tell you much on its own.

If your support volume is a phone queue, Retell or Vapi are the right category to be shopping in. If your support volume is the ticket queue sitting inside your help desk, a phone-first voice platform is the wrong tool even if it's a well-built one: none of Retell's pricing or features touch the inbox where most support teams actually spend their day. That's the gap Macha is built for, an AI agent layer that reads and answers inside Zendesk, Freshdesk, Gorgias, Front, HubSpot or Intercom rather than a separate voice stack you have to wire up and monitor yourself. Macha is also done for you: the Macha team analyzes your past tickets, builds the knowledge base and the agent, and runs it in safe mode until the drafts are right, included at about $0.40 a ticket with a dedicated success manager who handles your changes. For a broader look at how AI voice platforms differ from ticket-based agents, see our guide to AI agents for customer service; for another enterprise voice comparison, Parloa; and for a builder that runs voice and chat off the same agent with a credit meter instead of a per-minute one, our Voiceflow guide.

How we researched this

We fetched Retell's homepage, pricing page, and build/analytics documentation directly on 2026-09-17, along with Vapi's pricing page for the comparison, and re-checked both pricing pages, the Trustpilot score and Retell's billing-exceptions doc on 2026-09-24. G2's reviews page returned a 403 to automated access on that date; we read it in a browser on 2026-09-18 (4.8 from 2,638) and took the quotable reviews from Trustpilot's public page. For the problems section we read Retell's public support forum at community.retellai.com and its billing-exceptions, support and international-calling docs on 2026-09-20, and every thread cited there is linked with its date. We did not create a Retell account for this pass, so there is no first-hand call, no measured latency and no screenshot of our own; the numbers reported as measurements are the forum posters' own. Revenue and funding figures are press-reported or analyst estimates (TechCrunch, Sacra), not vendor-published, and are hedged accordingly. Customer results (Pine Park Health, SWTCH, Medical Data Systems) are vendor case studies we could not independently verify.

FAQ

What is Retell AI? Retell AI is a developer-focused platform for building AI voice agents that handle inbound and outbound phone calls, plus embedded chat agents, with per-minute pricing and component-level control over the model, voice, and transcription vendor.

How much does Retell AI cost? Pay-as-you-go pricing runs $0.07 to $0.31 a minute blended, starting with $10 in free credits and 20 free concurrent calls. The $0.07 floor covers only Retell's voice infrastructure and platform text-to-speech; add the cheapest LLM and telephony and a real call starts around $0.087 a minute. The final rate depends mostly on the LLM tier you choose, since that line item alone ranges from $0.0016 to $0.64 a minute. Enterprise pricing is custom.

Is there a free trial? Retell gives new accounts $10 in free usage credits and 20 free concurrent calls, with no fixed-length trial window to work against.

What has Retell AI's growth actually been? Sacra estimates Retell at about $80 million in annualized revenue as of August 2026, up from roughly $45 million at the end of 2025, on about $5.1 million raised since its Y Combinator start. Retell does not publish these figures itself.

Does Retell AI replace a help desk? No. Retell handles phone calls and embedded chat. It has no ticket queue or email channel, so teams running most of their support through a help desk would need a different tool, or one layered on top of the help desk itself, alongside it.

What do real users say about Retell AI? Trustpilot shows a 4.9/5 rating from 814 reviews as of 2026-09-24, 94% of them five-star and 1% one-star. Reviewers consistently praise how quickly they can build an agent, the pace of platform updates, and uptime; the one-star tail is almost entirely one complaint, no clear support channel after payment, raised by separate reviewers months apart.

How does Retell AI compare to Vapi? Both are developer-first voice AI platforms with per-minute, component-based pricing. Vapi adds a flat $29/month Core tier for small teams and a $999/month-minimum Pro tier; Retell stays fully usage-based until you reach its custom Enterprise plan. Run your own call-volume numbers through both pricing calculators; a headline rate alone won't tell you much.

What compliance certifications does Retell AI have? HIPAA, SOC 2 Type II, GDPR, and ISO 27001, with an enterprise on-premises deployment option, which is why healthcare and insurance appear repeatedly among its named customers.

Why is my Retell AI bill higher than the per-minute rate? Usually the long-prompt surcharge. Retell's billing-exceptions doc says a prompt over 4,000 LLM tokens scales the billed duration of the whole call, using Scaling Factor = Prompt LLM Tokens / 4,000 and Billed Duration = Original Duration x Scaling Factor, rounded up. The token count includes the global prompt, function descriptions, node prompts, the transcript so far, tool history and knowledge-base content, so a stable agent can drift over the line as conversations get longer. Calls under 10 seconds that use a dynamic opening message also carry a 10-second minimum.

Why do Retell AI call transfers fail? The error "Transfer call cannot be completed, telephony provider or carrier declined or was unavailable" points at the carrier. Retell's support threads give four fixes: switch the transfer mode from SIP REFER to SIP INVITE and confirm your trunk supports REFER at all, switch to warm transfer, use the agent's number as the caller ID instead of the caller's, or add custom SIP headers. For extensions, set optOutHumanDetection to true on the transfer_call tool and avoid a Displayed Caller ID of "User's Number".

Ready to automate the ticket queue instead of the phone queue? See how Macha's pricing works or start a trial on top of the help desk you already run.

Sources: Retell AI, Retell AI pricing, Retell build docs, Retell analytics dashboard docs, TechCrunch, "Retell AI lets companies build 'voice agents' to answer their calls", Trustpilot: Retell AI reviews, Vapi pricing, Sacra: Retell AI research.