business

Best AI Voice Agents

The best AI voice agent platforms, ranked on independent evidence: per-minute pricing, latency, voice quality, developer control and compliance, plus the speech-to-speech models underneath them, the benchmarks that test them, and the rules on AI phone calls.

Updated October 2026

Quick answer: The best AI voice agent platform for most businesses is Retell AI, because it has the largest body of independent user ratings in the category, 4.8 out of 5 from 2,638 reviews on G2, publishes itemised per-minute pricing (from about $0.07 a minute before telephony) and lets non-developers build agents in a visual flow editor. For the most natural-sounding agent with a free plan to start on, ElevenLabs Agents is the better choice, because it scored a 90.5% task success rate in the independent Artificial Analysis Speech Agent Arena, the highest of any off-the-shelf cascaded agent platform tested. For engineering teams that want to choose every component themselves, Vapi is the pick, because both independent expert reviews we found rate it the most flexible platform, and it charges a $0.05 per-minute platform fee with model costs passed through at cost. The one caveat: independent research shows voice agents still complete far fewer tasks than the same AI working in text, with τ-Voice measuring 31% to 51% task completion against 85% for a text model, so test any agent on your real calls before it talks to customers.

This page ranks platforms for building an AI voice agent: software that holds a spoken conversation on the phone or the web, understands what the caller wants and takes actions such as booking, updating a record or transferring the call. That is a different job from three pages next to it:

• Best AI voice generators ranks text-to-speech: turning written text into a spoken voice.
• Best AI voice clone ranks tools that copy a specific person’s voice.
• Best AI voice changers ranks real-time tools that transform your own voice as you speak.
• This page ranks the systems that combine a voice with a language model, a phone line and tools, so the agent can hold a conversation and act on it.

If you want a ready-made answering service for a small business rather than a platform to build on, that is a separate purchase with per-call pricing and no setup. For the wider agent landscape beyond voice, see best AI agents, and for the developer libraries used to build agents, see best AI agent frameworks.

Which one should you pick?

What a voice agent is, and how it works

A voice agent is software that listens to a caller, works out what they want, takes an action and replies in a spoken voice, in real time, with no person on the line.

Every voice agent is built in one of two ways:

Most platforms on this page are orchestration layers. Coval’s founder Brooke Hopkins put it plainly in a February 2026 comparison: Vapi and Retell “orchestrate existing providers (STT, LLM, TTS) into unified pipelines” and neither builds its own voice models. What you pay a platform for is the plumbing: phone numbers, interruption handling, tool calls, call transfer, recording, analytics and compliance.

A platform also decides how an agent connects to your systems. Most now let an agent call your own APIs, and several support the Model Context Protocol, the open standard for connecting AI to tools and data.

What the independent evidence says

Voice agents still lag text agents by a wide margin. τ-Voice, a benchmark published on 14 March 2026 by Soham Ray, Keshav Dhandhania, Victor Barres and Karthik Narasimhan, ran 278 customer-service tasks. A text-based GPT-5 completed 85% of them. Full-duplex voice agents completed 31% to 51% in clean audio and 26% to 38% with realistic noise and accents, keeping only 30% to 45% of the text model’s ability. The authors traced 79% to 90% of failures to the agent’s own behaviour.

No system is best at both accuracy and experience. ServiceNow AI Research’s EVA-Bench tested 12 commercially available voice agent systems, seven cascaded, two hybrid and three speech-to-speech. “No system simultaneously exceeds 0.5 on both EVA-A (Accuracy) and EVA-X (Experience)”: cascaded systems scored better on getting the task right, speech-to-speech systems on how the conversation felt.

People prefer Google’s live models, but preference is not task success. The Artificial Analysis Speech Agent Arena, announced on 24 August 2026, has people hold blind live conversations with two hidden agents on the same task, such as booking a dental appointment or ordering takeout, and vote for the better one. It also records whether the agent completed the task.

Speech Agent Arena: the top of the leaderboard

Source: Artificial Analysis. GPT-Realtime-1.5 anchors the scale at 1,000. Cascaded platforms were tested with their default pipelines rather than tuned set-ups.

Three readings matter for a buyer. First, the top-voted system, Gemini 3.1 Flash Live in its minimal setting, completed only 74.6% of tasks, while Grok Voice Think Fast 2.0 completed 94.6% with a lower preference score: the agent people liked talking to most was not the one that got the job done most often. Second, ElevenLabs Agents completed 90.5% of tasks, ahead of the other cascaded platforms tested, Cartesia Line at 77.5% and Deepgram Voice Agent at 73.7%. Third, Retell AI and Vapi are not on the leaderboard, so it cannot be used to compare them.

Vendor benchmarks are not evidence here. Retell AI states on its homepage that it achieved “95.7% workflow accuracy” benchmarked against six platforms; that is Retell’s own test. A GitHub benchmark of two phone-agent platforms from 26 September 2026 publishes its data under a Calldesk account and finds Calldesk winning 3 to 0, so we do not count it. Several “we tested every platform” listicles are published by voice agent vendors ranking themselves first, and one widely read comparison discloses that two of its listed tools “paid to be listed in this guide”.

How we ranked voice agent platforms

A platform is ranked here only when there is more than one independent assessment of it, from more than one kind of source: an independent benchmark, hands-on expert reviews, or user ratings with real volume behind them. Each company’s own claims are reported beside that evidence and labelled as its claims.

Two expert reviewers need a disclosure. Coval and Cekura both sell testing and monitoring tools that run on top of voice agent platforms, including Vapi and Retell. Neither sells a voice agent platform, which is why we count them as independent, but both have a commercial interest in the category as a whole.

We weighted user ratings by volume. Vapi’s 25 Product Hunt reviews and 3 G2 reviews are a thin user record, which is the main reason Vapi ranks third rather than higher.

Best AI voice agent platforms ranked

The best AI voice agent platforms, reviewed

1. Retell AI: best AI voice agent platform overall

Retell AI is the best voice agent platform for most businesses, because it combines the largest independent user record in the category with published per-minute pricing and a builder a non-developer can use.

The independent evidence. Retell AI holds 4.8 out of 5 from 2,638 reviews on G2, the largest body of user ratings for any platform on this page. G2 reviewers praise ease of use, cited in 494 reviews, and natural-sounding voices with low latency. The most common complaints are missing features such as international number support (116 reviews), the cost per call at scale (87 reviews) and a learning curve on advanced configurations (80 reviews). Coval’s February 2026 comparison recommends Retell for “visual flow design preference, non-technical team involvement, pricing transparency” and “clearer cost forecasting”, and puts its latency at 600 to 800 milliseconds and its realistic total cost at $0.11 to $0.31 a minute.

Pricing. Retell’s pricing page breaks every minute into parts: $0.055 a minute for Retell’s voice infrastructure, a language model from $0.0016 a minute (GPT-5 nano) to $0.32 a minute (GPT-6 Astra), text-to-speech at $0.015 to $0.040 a minute, and telephony from $0.015 a minute, depending on country. The first 20 concurrent calls are free; each extra concurrent call costs $8 a month. Add-ons include knowledge bases ($8 a month each after the first ten), advanced denoising ($0.005 a minute), safety guardrails ($0.005) and PII removal ($0.01). New accounts get $10 of free credit.

What Retell claims. Retell states it is HIPAA compliant, SOC 2 Type II certified and compliant with GDPR and ISO 27001, with latency of about 600 milliseconds (Retell AI).

Pros: largest independent user record; itemised pricing; free concurrency up to 20 calls; visual builder for business teams; compliance certifications for regulated industries.

Cons: not tested in the Speech Agent Arena; users report costs rising at scale; international phone numbers are a recurring complaint on G2.

Best for: a business that wants an agent live quickly and wants both operations staff and developers to work on it.

2. ElevenLabs Agents: best for voice quality and starting free

ElevenLabs Agents is the best choice when how the agent sounds matters most, because it completed 90.5% of tasks in the only independent arena that tests agent platforms in live conversation, and it is the only ranked platform with a free monthly allowance.

The independent evidence. In the Artificial Analysis Speech Agent Arena, ElevenLabs Agents in its default cascaded set-up ranked 9th of 26 systems on preference, with an Elo of 993 from 773 samples, and completed 90.5% of tasks, the highest task-success rate of the three cascaded platforms tested. ElevenLabs as a whole holds 4.5 out of 5 from 1,239 reviews on G2; 239 of those reviews mention voice quality and ease of use when building voice agents, while one reviewer said the interface “now feels like you need a tour” since agent features were added. That G2 score covers all of ElevenLabs, including text-to-speech, not Agents alone.

Pricing. The ElevenAgents pricing page includes minutes on every plan:

Extra call minutes cost $0.08 each on every plan, and burst minutes above your concurrency limit cost $0.16. The language model is billed separately by model choice; telephony is included at no extra charge.

What ElevenLabs claims. ElevenLabs renamed its Conversational AI product to ElevenLabs Agents on 3 September 2025 and said more than 2 million agents had been created on it (ElevenLabs). Its voices are reviewed on our best AI voice generators page.

Pros: strongest independent task-success result of any cascaded platform; free plan with real minutes; included minutes make monthly cost predictable; ElevenLabs’ own voice library.

Cons: language-model costs sit outside the per-minute price; no expert review of Agents specifically found; G2 users report the interface got more complex as features were added.

Best for: customer-facing agents where the voice is the brand, and anyone who wants to try a voice agent without paying.

3. Vapi: best for developers who want full control

Vapi is the best choice for engineering teams, because both independent expert reviews rate it the most flexible platform for choosing and swapping every model, voice and phone provider.

The independent evidence. Coval’s February 2026 comparison recommends Vapi for “engineering-heavy teams preferring code-first development, maximum customization, version-controlled configurations, and broad provider/language options”, and lists support for 100-plus languages against Retell’s 30-plus. Coval’s separate May 2026 review calls it a “strong fit for teams wanting an open, swap-friendly voice AI stack” and puts cascaded latency at about 500 to 700 milliseconds. Cekura’s founding engineer Dileep Chagam, who tested Vapi directly, wrote on 9 May 2026 that it “excels for engineering teams needing full stack control” but that “when something breaks, you’re largely on your own unless you are on an enterprise plan”, with support limited to Discord below enterprise. Vapi holds 4.9 out of 5 from 25 reviews on Product Hunt, and 3.0 from 3 reviews on G2, where one healthcare reviewer reported the HIPAA add-on doubling from $1,000 to $2,000 a month without warning.

Pricing. Vapi’s pricing page charges a $0.05 a minute hosting fee, with transcription, language model and voice providers passed through at cost. Vapi’s own example puts 1,000 minutes a month at $82 to $129. Cekura measured real set-ups at $0.07 to $0.25 a minute. Optional success packages run from Core at $29 a month (10 concurrent calls) to Pro at 10% of the hosting fee with a $999 monthly minimum (30 concurrent calls). HIPAA compliance is a $2,000 a month add-on.

What Vapi claims. Vapi raised a $50 million Series B led by Peak XV on 12 May 2026, bringing total funding to $72 million, and said it supports more than 1 million developers, 2.7 million deployed agents and more than 1 billion calls (SiliconANGLE).

Pros: the widest choice of providers; lowest platform fee on this page; strong developer tooling; large developer base.

Cons: thin user-rating record; costs stack across several providers; support is limited below enterprise; HIPAA costs $2,000 a month.

Best for: engineering teams building a voice product, or companies with specific model, voice or telephony requirements.

The models underneath: speech-to-speech APIs

If you have engineers and want the lowest cost or lowest latency, you can build directly on a speech-to-speech model instead of a platform. You then provide the phone system, tool calls, monitoring and compliance yourself. These are models, not platforms, so they are not ranked above.

The arena shows why platforms still exist. The cheapest model on paper, Gemini Live, comes with no phone numbers, no call transfer, no dashboard and no compliance paperwork, and the most-liked model in conversation was not the most reliable at finishing tasks.

What a voice agent really costs

Per-minute prices on this page are not directly comparable, because each platform includes different parts. The table shows what each price covers.

A worked example: 1,000 minutes a month. Vapi’s own estimate for 1,000 minutes is $82 to $129. Retell’s price list gives about $72 before telephony with its cheapest language model, and more than $400 with its most expensive. ElevenLabs’ Pro plan at $99 a month includes 1,238 minutes and telephony, with the language model added on top. These figures cover the platform only, not the time to build, test and monitor the agent.

The language model is the biggest swing factor. On Retell’s price list, the language model alone ranges from $0.0016 a minute to $0.32 a minute. Choosing a smaller model for routine calls is the fastest way to cut cost; choosing a speech-to-speech model is the fastest way to raise it.

Best AI voice agent for each job

Best for inbound customer service

Retell AI is the best pick for inbound support lines, because its visual builder lets the support team who knows the call flows maintain the agent, and its 20 free concurrent calls cover most small and mid-sized lines.

Best AI phone agent for a small team without developers

ElevenLabs Agents is the best pick, because its Free and Starter plans include minutes, so a small team can test a phone agent before committing money. If you want a finished answering service rather than something to build, buy one of those instead.

Best AI sales agent for outbound calls

Retell AI is the best ranked pick for outbound sales calling, because its per-call cost is transparent at volume and its concurrency can be raised at $8 per extra line. Bland AI is built for outbound campaigns and caps the Build plan at 2,000 calls a day, but it is not ranked (see below). Before any outbound AI calling in the United States, read the rules section: AI-generated voice calls to consumers need prior express consent. For AI tools across the whole sales process, see best AI for sales.

Best for developers building a voice product

Vapi is the best pick, because it lets you swap transcription, language model, voice and telephony providers independently, and both independent expert reviews rate it the most flexible.

Best for healthcare

Retell AI is the best ranked pick for healthcare, because it states HIPAA compliance and SOC 2 Type II certification without a separate fee on its pricing page. Vapi offers HIPAA compliance as a $2,000 a month add-on. Verify any HIPAA claim with a signed business associate agreement, not a marketing page.

Best for the lowest latency

A speech-to-speech model such as Google’s Gemini Live or OpenAI’s GPT-Live-1 is the pick if latency matters above all, because speech-to-speech systems scored best on conversational experience in EVA-Bench. Expect to build the phone and tool layers yourself.

Best for many languages

Vapi is the best pick for multilingual agents, because Coval lists 100-plus supported languages for Vapi against 30-plus for Retell.

Other voice agent platforms worth knowing (not ranked)

These platforms are real and available, but we did not find enough independent assessment, from more than one kind of source, to rank them.

Synthflow is a no-code voice agent builder with 4.5 out of 5 from 1,015 reviews on G2, where 145 reviews cite pricing as a concern and users report phone numbers available directly only in the US, Canada and Australia. Its pricing page now lists only enterprise contracts, which it says “start at $30,000 annually”. The hands-on reviews we found carry affiliate disclosures, so we did not count them.

Bland AI is an all-inclusive voice agent platform priced at $0.14 a minute on its Start plan and $0.12 a minute on its Build plan ($299 a month), with “no token charges” and no model pass-throughs, according to Bland. It caps the Start plan at 100 calls a day and 10 concurrent calls. It holds 5.0 from 11 reviews on G2, too few to rank on.

PolyAI builds enterprise voice agents for contact centres as a managed service rather than a self-serve platform. It holds 4.5 out of 5 from 33 ratings on Gartner Peer Insights. Pricing is not public.

Cartesia Line and Deepgram Voice Agent are cascaded agent stacks from speech-model companies. Both appear in the Speech Agent Arena, Cartesia Line at 6th with 77.5% task success and Deepgram Voice Agent at 18th with 73.7%, but we found no expert reviews or user ratings at volume for either as an agent platform.

The rules on AI phone calls

This section is general information, not legal advice. Check the rules for your own jurisdiction and industry with a lawyer.

Outbound AI calls in the United States need consent. In February 2024 the FCC unanimously adopted a Declaratory Ruling recognising that calls made with AI-generated voices are “artificial” under the Telephone Consumer Protection Act, which requires prior express consent for artificial or prerecorded voice calls to consumers. The ruling governs calls placed to consumers, so it matters most for outbound sales agents and AI callbacks.

Utah requires an honest answer when a caller asks. The Utah Artificial Intelligence Policy Act, as amended effective 7 May 2025, requires a business to disclose that a consumer is talking to AI when the consumer makes a “clear and unambiguous request” to know, with proactive disclosure required for higher-risk uses; the state can impose $2,500 per violation (Davis Polk).

California’s bot law covers online communication. Business and Professions Code §17941 prohibits using a bot “online” to mislead a person about its artificial identity to drive a sale or influence a vote. Its text is confined to online communication and requires intent to mislead.

EU callers must be told they are talking to AI. EU AI Act Article 50, enforceable from 2 August 2026, requires AI systems that interact directly with people to be designed so those people are informed they are dealing with AI, unless that is obvious.

Call recording consent is set by state law. Most US states require one party’s consent to record a call, and several, including California, require every party’s. Every platform on this page records and transcribes calls, so the agent’s opening line needs to cover recording for the states your callers are in.

The safe default is the same everywhere: configure the agent to say it is an AI when asked, and test that it does.

How to choose a voice agent platform

1. Decide who will maintain the agent. If operations staff will edit call flows, favour a visual builder (Retell AI). If engineers own it, favour an API-first platform (Vapi).
2. Test on your own calls. τ-Voice found voice agents complete far fewer tasks than text agents, and noise and accents cut completion further. Run your real call scripts, with real background noise, before launch.
3. Price a full month, not a minute. Add the language model, voice, telephony, concurrency and compliance add-ons. The cheapest platform fee can produce the most expensive bill.
4. Plan the hand-off to a human. Every agent will meet calls it cannot handle. Check how it transfers a call and what context it passes on.
5. Check compliance in writing. For healthcare, finance or EU callers, ask for the signed agreement, not the badge on the homepage.
6. Measure after launch. Track task completion, transfers and complaints per week. Testing tools such as Coval and Cekura exist because agents that pass a demo often fail on edge cases in production.

Frequently asked questions

What is the best AI voice agent?

The best AI voice agent platform for most businesses is Retell AI, which holds 4.8 out of 5 from 2,638 G2 reviews and publishes itemised pricing from about $0.07 a minute before telephony. ElevenLabs Agents is the best choice for voice quality and has a free plan, and Vapi is the best choice for developer teams that want to choose every component themselves.

What is an AI voice agent?

An AI voice agent is software that holds a spoken conversation with a person, usually on the phone, understands what they want and takes actions such as booking an appointment, updating a record or transferring the call. It combines speech recognition, a language model and a synthetic voice, or a single speech-to-speech model, with connections to a phone line and business systems.

How much does an AI voice agent cost?

An AI voice agent platform typically costs about $0.07 to $0.31 a minute all-in, going by independent estimates of real set-ups on Vapi and Retell AI. Vapi charges a $0.05 a minute platform fee plus provider costs, and estimates $82 to $129 for 1,000 minutes. Retell AI starts at about $0.07 a minute before telephony, rising mainly with the language model chosen. ElevenLabs Agents includes 1,238 minutes in its $99 a month Pro plan and charges $0.08 for each extra minute.

Is there a free AI voice agent?

Yes. ElevenLabs Agents includes 15 call minutes a month on its free plan, Retell AI gives new accounts $10 of credit, Vapi gives $5 of credit, and Google’s Gemini Live API has a free tier for developers. None of these free options is enough for a production phone line.

What is the difference between an AI voice agent and an AI voice generator?

An AI voice generator turns written text into spoken audio, for voice-overs or narration. An AI voice agent holds a two-way conversation in real time, deciding what to say and taking actions as it goes. Most voice agents use a voice generator as one component, alongside speech recognition and a language model.

What is an AI phone agent?

An AI phone agent is an AI voice agent connected to a phone number, so it can answer or place telephone calls. Platforms such as Retell AI, ElevenLabs Agents and Vapi provide or connect phone numbers, handle call transfer and record calls, so the same agent can run on a business phone line.

Can AI voice agents make outbound sales calls?

Yes. Retell AI, Vapi, ElevenLabs Agents and Bland AI all support outbound calling. In the United States, the FCC ruled in February 2024 that AI-generated voices are “artificial” under the Telephone Consumer Protection Act, so outbound AI calls to consumers need prior express consent.

Are AI voice agents reliable?

AI voice agents are reliable on narrow, well-defined tasks and much weaker on complex ones. The τ-Voice benchmark (March 2026) found voice agents completed 31% to 51% of customer-service tasks in clean audio, against 85% for a text model, and 26% to 38% with realistic noise and accents. In the Artificial Analysis Speech Agent Arena, task success ranged from 29.1% to 94.6% across systems.

What is the difference between Retell AI and Vapi?

Retell AI suits teams that want a visual flow builder, transparent all-in pricing and non-technical staff involved, while Vapi suits engineering teams that want code-first control and the widest choice of model, voice and telephony providers. Coval’s February 2026 comparison measured similar latency for both, 600 to 800 milliseconds for Retell and 550 to 800 milliseconds for Vapi.

Do I have to tell callers they are talking to an AI?

It depends on where your callers are. Utah requires a business to disclose AI when a caller clearly asks, and EU AI Act Article 50, enforceable from 2 August 2026, requires people to be informed they are dealing with AI unless it is obvious. The safe default everywhere is to configure the agent to say it is an AI when asked. This is general information, not legal advice.

Which AI voice model sounds the most natural?

In the Artificial Analysis Speech Agent Arena, where people vote blind in live conversations, Google’s Gemini 3.1 Flash Live and Gemini 3.8 Live ranked first and second, followed by OpenAI’s GPT-Live-1. Among off-the-shelf agent platforms tested in their default set-ups, Cartesia Line scored the highest preference and ElevenLabs Agents the highest task success.

Is Synthflow still a good choice?

Synthflow holds 4.5 out of 5 from 1,015 G2 reviews, with pricing the most common complaint. Its pricing page now lists only enterprise contracts starting at $30,000 a year, so it suits large deployments rather than small teams testing a first agent. We did not rank it because we found no independent expert assessment without an affiliate relationship.