The best AI voice agent platforms, compared on what actually costs

The best artificial intelligence voice agent platform depends on two things, how much assembly you tolerate and where you accept your calls to travel. To move fast, Retell AI is the shortest path between an idea and a phone number that answers. To control everything, Vapi lets you swap every brick, and LiveKit lets you host the whole thing yourself. For a French company that has to answer a data protection officer, Rounded and Zaion are the only two simple answers.

This article compares nine platforms on what they give the reader. Prices were taken from the vendors' own pages in August 2026, and this category changes its pricing every quarter.

How this comparison was built

Criteria come before the first name, otherwise a ranking is just a preference in disguise. Five questions settled it, in this order.

• Where the calls travel. Hosting in France or in the European Union changes the conversation with a legal team, and it is the first filter for a French company.
• What a minute really costs. Almost every platform advertises the price of its orchestration layer alone. The rest of the bill arrives later.
• How much control you keep on the stack. Some let you pick the transcription, the model and the voice. Others decide for you, which is faster and locks you in harder.
• What the platform can do during the call. An agent that only talks is nearly useless. Value comes from transfers, text messages, slot bookings and calls into your own systems.
• What the tool cannot do. Concurrent call limits, language coverage and the pricing floor eliminate more platforms than the advertised features do.

1. Vapi, open infrastructure brick by brick

Vapi decides nothing for you. You pick the transcription, the language model and the synthetic voice, and you can bring your own API keys. It is the platform that behaves most like infrastructure, with everything that implies in freedom and in work.

Key features

• You compose the transcription, model and voice chain freely, vendor by vendor.
• The agent calls your tools mid conversation, sends a text message, books a slot or transfers to a human.
• Outbound calls fire from a simple network request, which makes campaign launches trivial.
• The dashboard exposes full transcripts, duration and cost for every call.
• A zero data retention option exists, billed separately.

Underlying model and limits

No model is imposed, which is Vapi's main strength and its main burden. You own the trade-offs between latency, quality and cost, and each brick is tuned separately. Hosting is American, and stronger compliance comes through paid add-ons rather than the base plan.

Pricing

The Build plan is usage based, with no subscription. Vapi hosting is billed at 0.05 dollar per minute, roughly 4 euro cents, on top of which transcription, model and voice are charged at cost. Ten concurrent lines are included, then each line costs about 9 euros a month. HIPAA compliance runs around 1,700 euros a month and zero retention around 850. The Scale plan moves to an annual contract, as recorded in August 2026.

What users say

Vapi's G2 listing showed 4.2 out of 5 in August 2026, on three reviews only. That is far too thin to conclude anything, and saying so beats padding the section. The useful signal sits elsewhere, in the volume of tutorials and projects published around the platform.

My take on Vapi

This is the platform I built my first demo voice agent on, a switchboard that screens calls, sends a text message and transfers to a human. What I keep from it fits in one sentence, Vapi's freedom turns into a trap if you over equip the agent. Past three or four resolution paths it loses the thread and you have to fall back on a human transfer. I happily pick it up again when the team opposite can read technical documentation, and I rule it out when nobody wants to arbitrate between four voice vendors.

2. Retell AI, the shortest path to production

Retell AI absorbs the real-time plumbing and leaves you the script. You build in a visual editor where each node carries its own role, rules and tools. It is the platform that demands the fewest technical decisions before your first working call.

Key features

• A drag and drop flow builder replaces the single giant prompt with a business decision tree.
• Agents cover more than thirty languages, French included, with latency advertised around 600 milliseconds.
• The voice channel is doubled by a text and a chat channel, driven by the same agent.
• Testing tools simulate conversations and compare two versions of the same flow.
• Post-call analysis surfaces satisfaction, sentiment and conversation outcome.

Underlying model and limits

The model is picked from a list, from cheapest to most capable, and the per-minute price follows that choice. Hosting is American, with personal data removal options billed per minute. The price of the comfort is clear, you live inside the Retell world and migrations hurt.

Pricing

Pay as you go starts with 10 dollars of free credits. Retell bills its voice infrastructure at 0.055 dollar per minute, roughly 5 euro cents, speech synthesis from 0.015 to 0.040 dollar, and the model from 0.003 dollar for the smallest to 0.16 for the most capable. The advertised range runs from 0.07 to 0.31 dollar per minute, roughly 6 to 27 euro cents. Knowledge base and personal data removal add on top, as recorded in August 2026.

What users say

This is the best documented platform by its own users, at 4.8 out of 5 across more than 2,600 G2 reviews in August 2026. Ease of use and the interface come up most often as praise. Two complaints stand out clearly, missing features on complex scenarios and a cost judged high once volume grows.

My take on Retell AI

I have not deployed it on an engagement, so I stick to what documentation and reviews allow me to verify. On paper it is the obvious answer for a team that wants an agent answering the phone next week with no engineer to spare. The point I would watch is cost at scale, because that complaint recurs too often to be anecdotal.

3. Rounded, the French alternative hosted in France

Rounded is a French platform that does roughly what Vapi does, in French and from France. It started in 2023 and targets organisations that cannot send their calls across the Atlantic. Clinics and hospital groups feature prominently in its public references.

Key features

• Servers carry the French health data hosting certification, which unlocks medical use cases in France.
• The catalogue advertises around two thousand native integrations, including the main customer relationship tools.
• Inbound and outbound calls run from the same interface, triggered by a network request.
• More than thirty languages are covered, with response time advertised around 700 milliseconds.
• Tiers scale up to more than two hundred concurrent calls on the custom plan.

Underlying model and limits

Rounded assembles the same bricks as its American competitors, transcription, language model and speech synthesis, but orchestrates from French infrastructure. The limit is that of a mid-sized vendor facing platforms funded in the hundreds of millions. Public documentation is thinner, and the developer community far smaller.

Pricing

It is the only platform in this comparison priced entirely in euros. The trial includes 15 minutes, the Essential tier costs 129 euros a month for 500 minutes then 0.20 euro per minute, the Pro tier 489 euros for 2,000 minutes then 0.15, and the Plus tier 1,189 euros for 7,000 minutes then 0.13. The custom plan goes down to 0.07 euro per minute. A one-off pack of 100 minutes exists at 19 euros, as recorded in August 2026.

What users say

No software marketplace shows a usable review volume for Rounded in August 2026. An invented figure would be worse than none, so the only public signal available remains the reference list the vendor displays, which includes French clinics and software companies.

My take on Rounded

This is the platform I use for outbound calls. The setup that works is simple, a list in a spreadsheet, a trigger, and a network request passing first name, last name and due date as variables. The agent then opens with the person's name, which completely changes the hook compared to a classic call centre call. I recommend it without reservation as soon as hosting becomes a question, and I switch to Vapi when I need an unusual voice vendor.

4. ElevenLabs Agents, when the voice carries the value

ElevenLabs built its reputation on speech synthesis before turning it into an agent platform. The result is coherent, this is the offer where the voice itself carries the value. The vendor lets you steer intonation directly inside the text, with cues that change how something is said without changing what is said.

Key features

• Voice quality remains the market reference, in English as in French.
• More than seventy languages are covered by the same platform.
• Intonation is driven by cues placed in the text, which makes emotion controllable.
• Agents work in voice and in text, over the phone and on the web.
• The free tier allows fifteen minutes of calls with no card required.

Underlying model and limits

The voice is in house, the rest of the chain leans on third-party models. The most annoying limit for a European company is administrative, European data residency is reserved for enterprise customers. Below that, you stay on the default infrastructure. Concurrent calls are also capped per tier, from four on the free plan to forty on the highest one.

Pricing

The grid is a classic subscription with included minutes. The free tier gives 15 minutes, Starter 6 dollars for 75 minutes, Creator 22 dollars for 275 minutes, Pro 99 dollars for 1,238 minutes, Scale 299 dollars for 3,738 minutes and Business 990 dollars for 12,375 minutes. Extra minutes cost 0.080 dollar, roughly 7 euro cents. Recorded in August 2026.

What users say

The ElevenLabs G2 listing covers the whole vendor, from voice cloning to dubbing, not the agents product in particular. It therefore says almost nothing about the product compared here, and writing that down beats presenting it as a verdict on the agent platform.

My take on ElevenLabs Agents

I have not deployed their agent platform on an engagement, but I have used their voices in audio builds for a long time. What I observe is consistent, the naturalness gap between English and French is still real, even there, and it narrows every quarter. I would pick them without hesitation on a case where the agent represents the brand and where a robotic voice would cost dearly in perception.

5. Bland AI, high outbound volume at a predictable price

Bland AI was designed for volume, with per-minute pricing that does not move with the model used. In August 2026 the platform tightened its message around healthcare and high-stakes calls, where reliability beats flexibility. Its public counter had passed 630 million resolved calls by then.

Key features

• Journeys are drawn as deterministic trees, which keeps behaviour predictable across thousands of calls.
• The per-minute price is fixed per tier, with no surprise tied to the chosen model.
• Tiers scale to a hundred concurrent calls and five thousand calls a day outside enterprise plans.
• Transfers to a human are billed at a reduced rate, which matters on campaigns that escalate often.
• The entry tier includes an inbound number with no card required.

Underlying model and limits

The stack is integrated and the vendor does not publicly detail which model runs behind it. That is the price of pricing predictability. Hosting is American, and configuration flexibility stays below Vapi. The daily call cap on the first tiers catches out teams that launch a large campaign without reading the grid.

Pricing

Three public tiers and an enterprise plan. The Start tier is 0.14 dollar per minute with no fixed fee, the Build tier 299 dollars a month plus 0.12 per minute, the Scale tier 499 dollars a month plus 0.11. Transfer minutes cost 0.03 to 0.05 dollar. In euros, expect roughly 10 to 12 cents a minute, as recorded in August 2026.

What users say

The G2 listing showed 5 out of 5 in August 2026, on eleven reviews. The score is excellent and the sample is too small to be representative, both are true at once. Recurring praise covers speed and the flexibility of the programming interface.

My take on Bland AI

I have not deployed it on an engagement. What looks solid in their proposition is price stability, because that is exactly what is missing elsewhere when you have to budget a hundred thousand minutes. What I would check before signing is the daily cap on the chosen tier, and real French language quality, which stays the weak point of American platforms.

6. Synthflow, the no-code tool that went enterprise only

Synthflow used to be one of the most accessible no-code platforms on the market, with an entry tier at a few dozen euros a month. That is no longer true. In August 2026 its pricing page shows a single enterprise offer, and the old tiers are gone. Most comparisons published elsewhere still quote prices that no longer exist.

Key features

• Telephony is built in natively, with the option to plug an existing SIP trunk.
• Integrations cover sales software, calendars and contact centre stacks.
• Escalation and fallback paths are designed with the vendor during rollout.
• Launch support, training and continuous optimisation are part of the contract.
• The vendor claims more than 65 million calls handled every month across thirty countries.

Underlying model and limits

Synthflow does not publish the composition of its stack. The main limit is no longer technical, it is commercial. A small organisation simply cannot get in any more, and a team that started on an old tier has to redo its decision. That is a useful lesson for the whole category, today's accessible platform can shut its door tomorrow.

Pricing

One public offer only. Enterprise contracts start at 30,000 dollars a year, roughly 26,000 euros, with final pricing scoped on call volume, concurrency, telephony, integrations and security requirements. No self-service tier is displayed, as recorded in August 2026.

What users say

The vendor itself displays 4.5 out of 5 across more than a thousand G2 reviews on its pricing page, which makes it the second best documented listing in the category behind Retell. The recurring complaint in those reviews is limited customisation compared with developer-oriented platforms. The praise covers ease of adoption, exactly what has just disappeared from the grid.

My take on Synthflow

I have not deployed it on an engagement, and the pricing change makes the verdict easy to write. If you are looking for a no-code platform to test a voice agent, this is no longer the place. If you are a large account wanting contractual support through rollout, the offer makes sense, but it then competes with Parloa and Zaion, not with Vapi.

7. LiveKit Agents, the open source you host wherever you want

LiveKit is not a no-code platform, it is a real-time framework. The agents framework and the media server ship under an open licence, and you can run them on your own infrastructure. It is the only option in this comparison that makes you the owner of exactly where the audio travels.

Key features

• The framework and media server are open source and self-hostable with no licence fee.
• The managed service handles deployment, scaling and observation of agents.
• Telephony sits on top of real-time transport, with numbers and inbound minutes included per tier.
• Included inference credits let you call the most common models without a separate contract.
• Scale reaches six hundred concurrent sessions on the highest public tier.

Underlying model and limits

You plug in whichever models you want, and LiveKit handles the transport. The brick powers ChatGPT's advanced voice mode, which says a lot about its robustness. The vendor raised 100 million dollars in January 2026 at a one billion valuation, a round led by Index Ventures. The limit is plain, without an engineering team you will get nowhere. You have to deploy, monitor and fix.

Pricing

Self-hosting costs nothing in licence fees. The managed service offers a free Build tier with 1,000 session minutes, a Ship tier at 50 dollars a month for 5,000 minutes and a Scale tier at 500 dollars for 50,000 minutes. Beyond that, an agent session minute costs 0.01 dollar, under one euro cent, and an inbound telephony minute the same. Recorded in August 2026.

What users say

No software marketplace shows a usable review volume for LiveKit in August 2026, which is normal for a project whose users are engineering teams. The most reliable public signal is the size of the code repository and the names of the companies relying on it.

My take on LiveKit Agents

I have not put it into production on an engagement, and I still recommend it in one precise case. When a company absolutely refuses to let the audio leave its perimeter, it is the only honest answer in this comparison. The cost simply moves, it leaves the platform invoice and settles into your engineers' time.

8. Parloa, the European contact centre player

Parloa does not sell an agent builder, it sells an agent management platform for contact centres. The nuance matters, because it changes the whole vocabulary, supervision, quality control and escalation to human advisers. The company was founded in Berlin in 2018 and employs around three hundred people.

Key features

• Agents cover phone, chat, messaging apps and mobile applications from the same platform.
• Supervision resembles managing a human team, with quality control and metrics.
• Displayed certifications cover ISO 27001, SOC 2, PCI DSS and the European operational resilience regulation.
• Integration happens with existing contact centre stacks rather than against them.
• Public references include TUI, Decathlon, Generali, Booking.com, SwissLife and KPMG.

Underlying model and limits

The vendor does not publish its stack, and that is not what its customers buy. The limit is the entry ticket, invisible but real. Parloa addresses organisations that already run a structured contact centre. The company raised 310 million euros in January 2026, a Series D valuing it around 2.5 billion euros, which explains its aggressive commercial presence across Europe.

Pricing

No public pricing. Parloa works exclusively on request, after a sales demo. For an organisation running a comparison, that means starting a sales cycle before getting any order of magnitude, as recorded in August 2026.

What users say

Software marketplaces show no usable review volume, which is common for offers sold only through a sales team. The case studies published by the vendor cover insurers and travel companies, and they remain vendor documents.

My take on Parloa

I have not deployed it, and I keep it in mind for one precise profile. A customer service leadership running several hundred advisers does not compare Parloa with Vapi. It compares it with its current contact centre vendor. That is an infrastructure decision, with the decision cycle that comes with it.

9. Zaion, the French sovereign option for regulated sectors

Zaion is the oldest player in this comparison, founded in 2017, well before the current wave. The company employs around seventy people and claims more than fifty customers across banking, insurance, real estate, telecoms and the public sector. Its main argument is not technology, it is location.

Key features

• AI models are designed by the vendor and hosted on sovereign infrastructure.
• The offer covers the agent that handles the call, assistance to the human adviser and conversation analytics.
• Call handling is split into specialised agents, from qualification to claim filing.
• Displayed compliance covers GDPR, the AI Act, ISO 27001 and the European operational resilience regulation.
• The vendor claims more than thirty million conversations handled.

Underlying model and limits

Zaion develops its own speech recognition, language and synthesis models, which is rare and expensive. The vendor claims hosting on sovereign infrastructure and encryption of conversations, transcripts and logs. The trade-off is classic, an in-house model struggles to match the best general purpose models on conversational finesse. You swap subtlety for control.

Pricing

No public pricing, sales on request after scoping. As with Parloa, the order of magnitude only appears after a sales cycle, which effectively rules out small organisations. Recorded in August 2026.

What users say

No usable public review volume on software marketplaces. The references named on the vendor's site include mutual insurers and consumer service groups, and those are commercial references, not independent reviews.

My take on Zaion

I have not deployed it, and I know when I would raise it. When the IT leadership of a bank or an insurer asks about sovereignty before performance, the discussion ends on a very short list. Zaion is on it, and that is a selection criterion most English-language comparisons ignore entirely.

What one minute of an AI voice agent really costs

A platform's headline price almost never covers more than orchestration. One call minute is paid five times over, and four of those lines never appear on the pricing page you just read.

The first line is telephony, and many forget it. A call goes through a carrier before it reaches the platform. I use Twilio on my side, and getting a French number requires a company registration extract and a twenty-four to forty-eight hour delay. I detailed the whole chain in my newsletter edition on voice agents, with the three technologies that follow one another behind it.

The three middle lines form the classic pipeline. Voice becomes text, text becomes an answer, the answer becomes voice again. That is the dominant architecture in production, and it stays more controllable than native voice-to-voice. We compared both approaches in detail, and the conclusion still holds in August 2026.

What the AI Act changed for your voice agent in August 2026

This is the point almost no comparison covers, and it has become binding. The European AI Act has applied since 2 August 2026, and its article 50 requires people to be informed when they interact with an AI system. In practice, your agent has to say so, and it has to say so at the start of the call.

That is neither hard nor expensive, it is one opening sentence. The difficulty sits elsewhere, with teams that deployed an agent in 2025 using a script written to fool the ear. Those scripts need rewriting, and nobody does it spontaneously.

The second constraint is older and less well known. The CNIL points out that voice is biometric data, and that processing it to recognise a person requires explicit consent. An agent that recognises a customer by their voice therefore falls under a different regime from one that merely understands what they say.

The third constraint concerns recording. The CNIL frames the listening and recording of calls, with information given to people, a limited retention period and an entry in the record of processing activities. A voice agent that archives its transcripts falls squarely inside that frame.

Should you buy a platform or build your own voice agent

The question comes up in every scoping session, and I settled it publicly on LinkedIn. On a voice agent the answer is almost always the same. Buy the platform, build what differentiates you. The voice brick is not your competitive advantage, your business logic is.

What genuinely gets built are the actions the agent triggers. Checking an availability, creating a task in your business software, scheduling a follow-up, escalating to the right person. That is where the difference lies between a switchboard that answers and an agent that works.

Technically, the missing brick is almost always the same. An automation tool picks up the agent's events and sends them where they belong, which we describe in our guide to n8n. Without it you will have bought a clever answering machine, and you will miss what a voice agent really automates.

Which AI voice agent platform to choose in 2026

If you want a one-line answer, take Retell AI to move fast and Rounded if your data has to stay in France. The first puts you in production within days. The second spares you a painful conversation with your legal team.

If you have an engineering team and a very specific need, Vapi lets you assemble every brick and LiveKit lets you host it all. If you run a contact centre of several hundred advisers, the real comparison is between Parloa and Zaion. It turns on sovereignty and integration rather than on the per-minute price.

And keep perspective. None of these nine platforms transforms an organisation on its own, they all pick up the same phone. Transformation starts when the call triggers an action inside your systems, which we scope in an AI audit and deploy on our autonomous agent engagements. For the deployment method itself, our voice agent guide remains the starting point.