best 9 ai voice agent platforms

retell publishes the component breakdown directly: voice infrastructure, stt, tts, llm and telephony each with their own sub-rate, summing into the quoted band. no other vendor here does that, and it means the advertised $0.07 to $0.31 is the closest thing to an honest all-in number in this ranking. the spread is almost entirely which llm you pick.

$10 of free credit, 20 concurrent calls included as standard, and an enterprise on-premise tier that deploys retell inside your own infrastructure — unusual for a platform this self-serve.

the add-ons are where the arithmetic escapes. knowledge base, advanced denoising and safety guardrails are $0.005 a minute each, pii removal $0.01, and ai quality assurance $0.10 after the first hundred minutes. turn on the ones a real deployment tends to need and you can add $0.02 to $0.13 a minute over the base — a plausible doubling that the headline band doesn't hint at.

the agents framework and the media server are both apache 2.0, confirmed from the repository licence file rather than the readme. self-host and there is no livekit fee at all; you pay your stt, llm and tts providers directly and run your own infrastructure. that is the same escape hatch llama.cpp gives local models, and no closed platform here can match it.

on livekit cloud the $0.01 session fee is only orchestration, but livekit publishes explicit rate ranges for every other component, so a realistic total is arithmetic rather than guesswork: about $0.095 a minute for a mid-tier stack, as low as $0.03 with cheap components, over $0.25 with premium tts. the free tier gives 1,000 agent session minutes a month plus a us phone number.

the drawback is that no blended number exists anywhere. five separately metered lines is more transparent than vapi's silence and more work than retell's band, and concurrency on the free tier is only five simultaneous sessions.

pipecat is bsd 2-clause, confirmed from the repository, and fully self-hostable independent of daily's hosted product. it is a framework rather than a platform: you wire in whichever stt, llm and tts you want and pay those vendors directly. run it on your own servers and the platform cost is zero.

on pipecat cloud, daily charges $0.01 to $0.03 a minute for agent compute depending on tier, with idle capacity as low as $0.0005. one-to-one voice transport is free, krisp noise removal is free for the first 10,000 session minutes a month, and concurrency is advertised as unlimited on pay-per-use. a realistic hosted stack lands around $0.07 to $0.18 a minute.

the honest cost is effort. there is no blended price to compare, because assembling a total means pricing four or five separate vendors yourself, and there is no free-minute allowance on pipecat cloud compute. for a team that wants to be told what a call costs, this is the wrong end of the market.

'no token charges' is the useful claim and it holds: the quoted talk-time rate covers the llm, speech-to-text and text-to-speech, so the usual four-way arithmetic collapses to two. add telephony pass-through and a realistic call is roughly $0.12 to $0.17 a minute, which is competitive and easy to forecast.

concurrency runs 10, 50 and 100 across the three plans, the highest published ceilings here, and self-hosted or on-premise deployment is available for enterprise with us, eu and apac data residency on request. bland claims 400ms response against a 1,240ms industry average — its own figure and its own comparison.

two carve-outs. telephony sits outside the bundle unless you bring your own twilio or sip trunk, in which case bland says transfer fees disappear. and while bland asserts your data never passes through third parties, that is a data-isolation claim rather than a statement about whether it trains on your calls, which it doesn't address.

the $0.05 is purely vapi's fee. stt, llm and tts pass through at the provider's own rate — or at zero if you bring your own api keys, which is the genuinely clever part of the model — and telephony is charged by the provider rather than by vapi. for a team that already buys deepgram and elevenlabs at volume, that structure is close to ideal.

for everyone else it is opaque. vapi itemises the split between its fee and 'model provider cost' but never sums an example call, so we built the estimate ourselves: roughly $0.15 to $0.30 a minute for a typical stack. that figure is ours, not vapi's, and the fact that we had to construct it is the criticism.

the build plan includes 10 concurrent lines with more at $10 a line a month, and vapi claims sub-600ms response times. its trust and security page returned 404 when we looked for a statement on whether call recordings train its models.

the technical proposition is real: telnyx co-locates gpus with its telephony points of presence, eliminating inter-provider network hops, and claims sub-200ms latency as a result — the lowest figure any vendor here publishes, and its own claim.

the pricing language is the problem. $0.05 a minute is labelled 'all-in' and covers speech-to-text, text-to-speech and orchestration plus a roster of included voices. call control is $0.002 a minute, webrtc another $0.002, sip trunking fees apply, and llm tokens are billed separately per million. premium voices bill per character on top. a realistic standard call is $0.07 to $0.10, before any premium voice.

that is not an unreasonable price. it is an unreasonable description of it, and the page contains both the word and its contradiction. elevenlabs voices are bring-your-own-key only, and concurrency limits aren't published.

elevenlabs' voices are the category benchmark, which is why its agent product gets recommended constantly. the credit system is public — tiers from free through $990 a month, with stt at 330 credits a minute and tts billed per character — and agent usage draws from that same shared pool.

what we could not find, across the pricing page, the conversational-ai docs overview and the product page, is any agent-specific per-minute figure or a credits-per-agent-minute conversion. of nine platforms this is the only one where a buyer cannot work out the rate from the vendor's own material at all.

one useful thing did surface: the docs note that 'zero retention modes are available for stricter data control', which implies retention is the default and stricter handling is opt-in. whether standard-retention call data trains elevenlabs' models isn't stated. concurrency figures aren't published either — the tiers list team seats, which is a different thing.

the product is real and active — live demo, working app, maintained changelog — and $0.02 a minute is the lowest headline figure of the nine. you choose your own llm from openai, mistral, llama or a custom endpoint, and your own voice provider from elevenlabs, playht, cartesia or a custom clone.

that architecture strongly suggests a pass-through model like vapi's, where $0.02 buys orchestration and everything else is extra. but we could not confirm it, because the dedicated pricing page returned 404 on both the www and non-www domains during this review, and the detailed breakdown millis references as living in its docs was not reachable.

so the honest position is that this is a teaser rate of unknown scope. by the pattern of every other pass-through platform here, a realistic all-in would land somewhere between $0.10 and $0.20 a minute — our inference from the category, not a figure millis has published. free tier, concurrency, latency and data handling are all likewise unstated.

the enterprise proposition is legible even though the price isn't: sla, native telephony or sip trunking, custom concurrency planning with routing, escalation and fallback logic, integrations, a security review and launch support, all scoped into a contract starting at $30,000 a year.

the self-serve side is where it falls down. you can build and test agents from templates at no cost, and costs begin only when an agent goes live — but the per-minute rate for a live agent is not published anywhere, for either tier. of nine vendors this is the only one with no rate card at all.

the enterprise page advertises 'data fine-tuning' — using past calls to train smarter agents — which reads as per-customer tuning on that customer's own history rather than a statement about training synthflow's underlying models. it doesn't clarify the distinction, and for a buyer handing over recorded calls that distinction is the whole question.