Best AI voice agents: pick by the job

Search "best AI voice agents" and you get a ranked ten that treats Vapi, PolyAI, ElevenLabs, and a Shopify answering service as interchangeable. They are not. Ranking them is like ranking a tractor, a bicycle, and a taxi because all three move.

The useful question is which job you have. This page splits the market into four jobs, then compares the developer platforms that actually publish prices, using their own pricing pages as of 22 September 2026. DialMCP sits in the last job, as the personal outbound option, not as a contact-center winner.

A red British telephone box on a village verge with a sign reading Coins Not Accepted Here, next to a red post box — Photo: Mike Cattell / Flickr (CC BY 2.0).

Four jobs, one search query

Build a voice agent. You or your engineers assemble speech-to-text, a model, text-to-speech, telephony, and turn-taking. Vapi, Retell, and Bland sell this. You still own the prompts, the tools, the numbers, and the compliance. The build vs buy map covers the speech stack itself.

Answer the business phone. A customer dials you and software talks. Most of the "small business" market lives here, and many of those products cannot place a call at all. That comparison is AI voice agents for small business, not this page.

Run a contact center. High-volume inbound, brand voice, multilingual disputes, on-prem options. PolyAI lives here. Public prices usually do not.

Let an existing AI assistant place one outbound call from your number. That is DialMCP. It is not a speech stack. The agent you already run gets four tools and a real PSTN call.

A fifth thing shows up on these lists and is not an agent: a speech model. ElevenLabs text-to-speech and OpenAI's GPT-Live-1 are mouths. ElevenLabs also sells ElevenAgents, which is a different product with a different price list. Mixing the two is how a $6-a-month starter plan ends up next to a $30,000 voice-AI contract.

What the published prices look like

Built from each vendor's own pricing page, read on 22 September 2026. Where a number is not published, the row says so rather than estimating.

Product	Job	What the headline rate covers	Published price
Vapi	Build	Hosting only. Speech, model, and telephony are extra (or $0 with your own keys)	$0.05/min hosting. 10 concurrent lines included, then $10 per line per month. HIPAA add-on $2,000/mo. Zero-data-retention add-on $1,000/mo
Retell	Build	Voice infra plus the models you pick. Range is their published all-in	$0.07 to $0.31 per minute. 20 concurrent calls included. Numbers $2/mo
Bland	Build / outbound volume	LLM, STT, and TTS. Telephony billed separately	$0.14/min with no platform fee, or $0.12/min with $299/mo, or $0.11/min with $499/mo
ElevenLabs Agents	Build, voice-led	Hosted minutes. LLM and telephony extra	$6/mo for 75 min up to $990/mo for 12,375 min. Extra minutes $0.08. Burst $0.16
Synthflow	Enterprise no-code	Quoted	Enterprise contracts start at $30,000 a year. No self-serve rate published
PolyAI	Enterprise contact center	Quoted	Not published
DialMCP	Personal outbound via MCP	The whole call, from your number	Free during launch. US and Canada. One concurrent call

The $0.05 figure that keeps appearing next to Vapi is the hosting line, not the call. Bland's own pricing FAQ is the clearest first-party version of this: it says Vapi lists $0.05/min and Retell about $0.07/min for infrastructure, then you add the model, speech, and carrier, and that a typical production stack lands roughly $0.13 to $0.30 on Vapi and $0.11 to $0.25 on Retell. Bland's $0.14 / $0.12 / $0.11 includes the LLM, STT, and TTS. Telephony is still extra, on your carrier or theirs at pass-through cost.

Vapi: best if you want to own the stack

Vapi is a developer platform. The Build plan is usage-based, includes 60-plus minutes, and charges $0.05 per minute for hosting. Speech-to-text, the language model, and text-to-speech are billed at cost, or $0 if you bring your own API keys. Concurrency starts at 10 lines and then costs $10 per extra line per month. Call history on Build is 14 days.

That is the honest pitch: you pick Deepgram or whoever for STT, whatever model you already use, Cartesia or ElevenLabs for the mouth, Twilio or Telnyx for the number. Vapi sits in the middle and bills a hosting fee. It is the right purchase if an engineer on your team is going to live in that stack.

Two add-ons matter more than the per-minute rate once you are in production. HIPAA is $2,000 a month on both Build and Scale. Zero data retention is $1,000 a month. SOC 2, SSO, and RBAC are listed on Scale, not on Build. Do not treat the $0.05 line as "HIPAA-ready voice AI for five cents."

Wrong for: a shop that wanted a receptionist this week and has nobody to maintain five vendor invoices. Wrong for: one person who wanted their existing assistant to ring a dentist.

Retell: best if you want a managed phone agent with a public all-in range

Retell's pricing page leads with "true pay as you go" and a published range of $0.07 to $0.31 per minute. Pay-as-you-go includes the full platform, 20 concurrent calls, analytics, transcripts, simulation testing, and webhooks. There is no monthly platform fee on that plan. Calls are billed to the nearest second.

The range is real because the components are itemized. Retell Voice Infra is $0.055 per minute. Their own voices, MiniMax, Fish, Cartesia, and OpenAI voices are $0.015; ElevenLabs voices are $0.040. LLM rates depend on the model: GPT 5.4 nano is $0.01 per minute, Claude 4.5 Sonnet is $0.08, GPT 5.5 is $0.16 on the standard tier. US Twilio telephony is $0.015 per minute; SIP trunking and custom telephony are listed as no charge on Retell's side. Their own example on the pricing page, with a mid-range LLM and a Retell voice on custom telephony, lands at $0.11 per minute.

The same page describes Retell as SOC 2 certified and HIPAA-ready on the self-serve product. Custom MSA/DPA, SSO, a dedicated server, and 24/7 support with a named portal sit on Enterprise.

Wrong for: teams that need to swap every layer of the stack every week. That is Vapi. Wrong for: high-volume outbound with a single bundled per-minute number. That is closer to Bland.

Bland: best if you want one bundled per-minute number at outbound volume

Bland's Start plan is $0.14 per minute, $0 platform fee, no card required, two credits and an inbound number they value at $15 a month. Build is $0.12 per minute plus $299 a month. Scale is $0.11 plus $499. Transfer minutes are cheaper than talk time: $0.05, $0.04, then $0.03. Concurrent calls go 10 / 50 / 100. Daily caps go 100 / 2,000 / 5,000.

The pitch on the same page is that there are no token charges and no model-provider pass-throughs. LLM, STT, and TTS are in the per-minute rate. Telephony is not: you bring Twilio or a SIP trunk, or you use Bland's Twilio at pass-through cost. BYOT customers do not pay transfer fees.

Compliance on the self-serve plans is thinner than the homepage lists imply. The pricing FAQ says SOC 2 Type I and Type II, HIPAA-eligible with a signed BAA, GDPR, and PCI DSS, and that BAA, SSO, and data residency are Enterprise.

Wrong for: a developer who wanted to bring their own model every Tuesday. Wrong for: a personal assistant placing three calls a day from a mobile number the callee already has.

ElevenLabs Agents: best if the voice is the product

ElevenLabs sells three catalogs on one domain, and listicles mash them together. The Creative / API plans are credits for text-to-speech, cloning, and dubbing. ElevenAgents is billed in call minutes, not that credit pool.

On the Agents price list: Free is 15 minutes and 4 concurrent calls. Starter is $6 for 75 minutes and 6 concurrent. Creator is $22 a month ($11 the first month) for 275 minutes and 10 concurrent. Pro is $99 for 1,238 minutes and 20 concurrent. Scale is $299 for 3,738 minutes and 30 concurrent. Business is $990 for 12,375 minutes and 40 concurrent. Extra minutes are $0.08. Burst minutes, when you exceed the concurrent cap, are $0.16. Text messages are $0.003. The LLM and any telephony are billed separately, on usage.

There is no limit on how many agents you create. You are limited by concurrent calls and the monthly minute pool. That is a voice company selling a hosted agent, not a phone company selling a voice. If the sound of the call is the product, start here. If the product is "my assistant should ring this number," you are buying the wrong catalog.

Synthflow and PolyAI: quoted, not ranked

Synthflow's pricing page publishes one number: enterprise contracts start at $30,000 annually, scoped around volume, concurrency, telephony, integrations, and launch support. Older posts still quote $29-a-month starter tiers. Those tiers are not on the page today. Treat any per-minute Synthflow figure you see in a 2026 roundup as stale unless it cites this page.

PolyAI is an enterprise dialog platform. The homepage lists SOC 2, HIPAA, GDPR, and PCI DSS as standard, a proprietary model (Raven) trained on more than a billion enterprise conversations, and two build paths (a non-technical builder and an ADK). There is no public price. It does not belong on a list next to a $6 starter plan.

If you need that class of product, you are in a sales cycle, not a pricing-table cycle.

DialMCP: best if the agent you already run needs to place the call

Most of the products above want you to build or buy a voice agent. DialMCP assumes you already have an agent (Claude, Codex, Cursor, ChatGPT connectors, Gamut) and that the missing piece is the phone.

It is a hosted remote MCP server at https://mcp.dialmcp.com/mcp. Connect it, verify your own US or Canadian number by SMS, and the agent gets four tools: place_call, get_call, end_call, list_calls. The callee sees your number. The call works through IVR and hold, hangs up on voicemail rather than leaving one, and comes back with a transcript, a recording link, and a structured outcome.

The limits are the product, not a footnote. Safety is enforced by the server: AI disclosure on every call, 8 a.m. to 9 p.m. in the destination's local time, one concurrent call, three per hour, ten per day, no more than two per day to the same number, a six-minute default and a ten-minute cap, and a platform-wide opt-out list. Telemarketing, cold outreach, robocalling, lead generation, and debt collection are prohibited. It is free during launch.

That is a terrible contact-center platform. It is the right object if the job is "ask Claude to call the vet." The MCP-native phone-call agents (DialMCP, VOYP, AgentCall, and the rest of that smaller set) are compared on best AI phone call agent. Which APIs can actually originate a cell-phone call is the voice AI API map.

Three mistakes the ranked lists keep making

Headline rate versus all-in. Vapi's $0.05 is hosting. Retell's $0.07 to $0.31 is the range after you pick a voice and a model. Bland's $0.14 includes the model and speech and still excludes the carrier. Convert to a cost per minute with the same three lines filled in, or you are not comparing prices.

Direction. Answering a call and placing a call are different products under different law. The small-business guide is the place that walks the inbound-only terms. This page is the developer-platform and personal-outbound split.

Whose number the callee sees. Inbound forwarding keeps the number you already advertise. Outbound caller ID is a setting, and spoofing is illegal, so a vendor that cannot display a number you control is placing calls that look like calls from a stranger. DialMCP's whole point is the opposite: the number is yours because you verified it.

How to pick

Engineers who want to swap STT, LLM, TTS, and telephony: Vapi.
A managed phone agent with a public all-in range and 20 concurrent calls on day one: Retell.
Bundled per-minute economics at outbound volume, and you will accept Bland's stack: Bland.
The voice is the product: ElevenLabs Agents.
A quoted enterprise contact-center or no-code rollout: PolyAI or Synthflow, through sales.
The assistant you already use should place a specific outbound call from your number: DialMCP.

If you came here for a single winner, that is the wrong deliverable. The lists that provide one are selling you their own row.

Want the assistant you already use to place the call? Connect DialMCP and let it ring a real number from your own verified line.

Connect DialMCP

FAQ

What is the best AI voice agent?

There isn't one. Vapi, Retell, and Bland are developer platforms for building a voice agent. PolyAI is an enterprise contact center. ElevenLabs Agents is a hosted agent from a speech company. DialMCP is a phone tool for an agent you already run. Pick the job, then the vendor.

How much does an AI voice agent cost?

Published self-serve rates on 22 September 2026 run from Vapi's $0.05/min hosting (plus speech, model, and telephony) to Bland's $0.11 to $0.14/min bundled, Retell's $0.07 to $0.31/min, and ElevenLabs Agents from $6/mo. Enterprise no-code starts at $30,000 a year at Synthflow. DialMCP is free during launch.

Is Vapi cheaper than Retell?

Not from the headline. Vapi's $0.05 is hosting only. Retell's published range already includes its voice infra and the models you pick. Bland's own FAQ estimates typical all-in production stacks at roughly $0.13 to $0.30 on Vapi and $0.11 to $0.25 on Retell. Add your actual STT, LLM, TTS, and carrier before you decide.

Can I use my own phone number?

For inbound, most receptionist products forward your existing number. For outbound, caller ID is a setting and you have to own the number you display. DialMCP uses your SMS-verified mobile as caller ID. Developer platforms typically provision a new DID or connect your Twilio account.

Do I need to build a voice agent to make an AI phone call?

No. If you already have an MCP client, you can connect a hosted phone server and skip the speech stack. That is the fourth job on this page. Building the stack is the right path when the voice itself is the product.

Best AI voice agents: pick by the job, not the demo.