Should I use a voice agent platform or assemble my own from APIs?

Platforms (Retell, Vapi, Bland) bundle speech-to-text, the LLM conversation, text-to-speech, and telephony into one integration — fastest to ship. Building blocks (Deepgram for speech-to-text) give you best-in-class components but you assemble and operate the pipeline yourself. LiveKit sits between: open-source realtime infrastructure and an agents framework, hosted on LiveKit Cloud or self-hosted.

Best Voice AI APIs for AI Agents

Q: What is the best voice AI API for AI agents?

For most teams, Retell AI: true pay-as-you-go with no platform fee and no monthly minimum. Its components are priced on one page — Voice Infra $0.055/minute, text-to-speech from $0.015/minute, plus per-minute LLM and telephony charges — totaling $0.07–$0.31/minute depending on the models you choose. Vapi offers more component-level control for developers, ElevenLabs has the best voice quality in the category, and LiveKit is the pick if you want an open-source stack you can self-host.

Q: How much does a voice agent call actually cost?

Budget from the all-in per-minute total, not the headline. Vapi charges a $0.05/minute hosting fee plus speech-to-text, LLM, and voice models at provider cost — its own pricing-page calculator example totals $82–$129 for 1,000 minutes ($0.082–$0.129/minute) with default providers. Retell's components total $0.07–$0.31/minute. ElevenLabs bundles agent minutes into its plans (15 minutes on Free up to 12,375 on Business) and charges $0.08/minute beyond the allowance — but LLM costs are billed separately on top, at per-model rates it does not publish. Bland bundles the LLM, transcription, and voices into $0.14/minute of talk time, with telephony billed separately. A five-minute call therefore lands roughly between $0.35 and $1.55 on Retell, about $0.41–$0.65 on a default Vapi stack, $0.40 plus LLM and carrier costs beyond your ElevenLabs plan minutes, and $0.70 plus telephony on Bland.

Q: Which voice AI API has the best free tier?

LiveKit: 1,000 free agent session minutes every month, no credit card required. Deepgram gives the largest one-time credit — $200, which its FAQ equates to roughly 43,000 minutes of Nova transcription. Retell gives $10 in signup credits with 20 free concurrent calls; Vapi gives $5 in credits with one phone number included; ElevenLabs' free plan includes 10,000 text-to-speech credits and 15 conversational-agent minutes a month.

Q: Can I self-host a voice agent stack?

Yes — LiveKit's server, SIP, egress, and Agents framework are open source (Apache 2.0); self-hosting removes the LiveKit platform fee and concurrency quotas, and you pay only your own infrastructure. Deepgram also offers self-hosted containers in your own VPC or on-prem at the enterprise tier. Vapi, Retell, and Bland are hosted services.

The best voice AI API for most agents is Retell AI: true pay-as-you-go with no platform fee — its components (Voice Infra $0.055/min, text-to-speech from $0.015/min, plus LLM and telephony) total $0.07–$0.31/minute depending on the models you pick, all itemized on one pricing page. Vapi gives developers more control, charging a $0.05/minute hosting fee plus models at provider cost. ElevenLabs — verified in full for this update (October 14, 2026) — brings the category's best voice quality, with agent minutes bundled into plans from free and $0.08/minute beyond the allowance.

Last checked: October 2026 · Researched from official documentation — hands-on API tests rolling out (methodology)

Best overall Retell AI
Best for developers Vapi
Best voice quality ElevenLabs
Best free tier LiveKit
Best speech-to-text Deepgram

#	Tool	What it is	Starting price (as published)	Free tier	Notable strength / limit
1	Retell AI	Platform to build, test, deploy, and monitor production AI voice and chat agents, with telephony built in.	No platform fee, no monthly minimum. Components: Voice Infra $0.055/min + TTS $0.015/min (ElevenLabs voices $0.040–$0.10/min) + LLM per minute + telephony $0.015/min → $0.07–$0.31/min total	Yes — $10 signup credits, full platform access, 20 concurrent calls included	Every cost component is itemized on one page — nothing is bundled into an opaque rate. Bring your own SIP trunk and Retell charges no telephony fee; Retell numbers are $2/mo.
2	Vapi	Developer platform for building, testing, and deploying voice agents, with bring-your-own transcriber, LLM, and voice models.	$0.05/min hosting fee; models passed through at provider cost, no markup. Pricing-page calculator example: 1,000 min = $82–$129/mo all-in with default providers	Yes — $5 free credits, 1 phone number included, 4 concurrent calls	Model-agnostic pass-through billing is genuinely flexible — but your total moves with every model choice. Watch the add-ons: HIPAA compliance is priced at $2,000/mo.
3	ElevenLabs	AI voice platform — best-in-class text-to-speech and voice cloning — with a Conversational AI agents product billed per call minute, separate from its TTS credit pool.	Agents: minutes bundled by plan — Free $0 (15 min/month, 4 concurrent calls), Starter $6/mo (75 min, 6 concurrent), Creator $22/mo (275 min, 10), Pro $99/mo (1,238 min, 20), Scale $299/mo (3,738 min, 30), Business $990/mo (12,375 min, 40). Beyond the allowance: $0.08/min (burst $0.16/min); text messages $0.003 each. TTS API, billed from the credit pool (Free 10,000 credits; Starter 30,000; Creator 121,000; Pro 600,000; Scale 1.8M; Business 6M): $0.08 per 1,000 characters on multilingual models, $0.04 per 1,000 on Flash/Turbo. Speech Engine API $0.08/min	Yes — Free plan: 10,000 TTS credits/month plus 15 Conversational AI minutes (4 concurrent calls)	The voice-quality benchmark — the reason its voices are resold inside rivals' pricing — and telephony carries no ElevenLabs fee on any plan (bring your own carrier, e.g. Twilio/SIP, billed at their rates). The limit is bill shape: agent minutes sit outside the credit pool, and LLM costs are billed separately on top at per-model rates ElevenLabs does not publish, so the true per-minute total needs its dashboard, not its pricing page.
4	LiveKit	Open-source framework and developer platform (LiveKit Cloud) for building, deploying, scaling, and observing voice, video, and physical AI agents.	Usage past the free tier: $0.01/min per agent session beyond included minutes, plus a plan fee on paid tiers (tier fees did not render on the pricing page during our check — unverified). Cloud calculator example: $0.0479/min	Yes — 1,000 free agent session minutes monthly, no credit card required	The only open-source, self-hostable option here (Apache 2.0): self-hosting removes the platform fee and concurrency quotas entirely. The trade-off is assembly and losing Cloud-only features like managed inference and observability.
5	Deepgram	Unified APIs for speech-to-text, text-to-speech, and voice-agent orchestration — a component layer, managed, in-VPC, or self-hosted.	Pay-as-you-go, no minimums. STT from $0.0043/min (Nova-3 pre-recorded); Voice Agent API $0.075/min standard ($0.050/min if you bring your own LLM and TTS); TTS from $0.0150 per 1,000 characters	Yes — $200 free credit (≈43,000 minutes of Nova transcription, per Deepgram's FAQ)	Transcription economics nobody else here matches, and its Voice Agent API unifies STT + LLM orchestration + TTS with discounts for bringing your own models. It's still a layer, not a phone system — telephony and agent logic are yours.
6	Bland AI	Enterprise voice AI platform for building, running, and monitoring AI phone agents for inbound and outbound calls.	Start: $0 platform fee + $0.14/min talk time ($0.05/min for transfers). Build: $299/mo platform fee + $0.12/min. Enterprise custom	Free to start — the Start plan includes 2 credits and an inbound number, capped at 100 calls/day	The per-minute rate bundles LLM, real-time transcription, TTS, and premium voices — no token charges, no provider pass-throughs. Telephony is separate: bring your own Twilio/SIP, or use Bland's built-in Twilio at pass-through cost.

Verdicts

1. Retell AI Best overall

The most legible pricing in the category: no platform fee, no minimum, every component priced on one page, telephony included or bring-your-own at no Retell charge. The $10 signup credit and 20 included concurrent calls make it the easiest serious test drive here. Your rate floats with your model choices ($0.07–$0.31/min), but it floats in the open.

2. Vapi Best for developers

The tinkerer's platform: swap transcriber, LLM, and voice per agent, all passed through at provider cost with no markup on top of the $0.05/min hosting fee — the official calculator puts a default 1,000-minute month at $82–$129 all-in. Two cautions from its own pricing page: the free credit is $5 (thinner than rivals), and HIPAA is a $2,000/mo add-on.

3. ElevenLabs Best voice quality

Verified in full on October 14, 2026 — and it takes the quality crown it was always going to take: its text-to-speech and voice cloning are the reference standard (Retell literally resells its voices at $0.040–$0.10/min). As an agent platform it is competitively priced: minutes are bundled into every plan (15 free, up to 12,375 on Business), extra minutes are $0.08, and telephony adds no ElevenLabs fee at all — you bring a carrier and pay their rates. It ranks third, not first, because the bill has three moving parts: the credit pool for TTS, per-minute agent billing outside that pool, and LLM costs billed separately at per-model rates that are not published on its pricing pages. Best voices in the category; a less legible invoice than Retell's.

4. LiveKit Best free tier

The infrastructure choice: 1,000 agent session minutes free every month, a published $0.01/min overage rate, and an Apache 2.0 core you can self-host to eliminate platform fees and concurrency quotas altogether. One transparency gap we hit ourselves: the paid tiers' monthly fees didn't render on the pricing page during our check, so we've left them marked unverified rather than quote them from elsewhere.

5. Deepgram Best speech-to-text

The component layer under half the stacks in this category. Nova-3 transcription at $0.0043/min pay-as-you-go, a $200 starter credit worth roughly 43,000 transcription minutes, and a Voice Agent API ($0.075/min, cheaper when you bring your own LLM/TTS) that closes much of the gap to a full platform. Deployable as a managed API, in your VPC, or self-hosted at enterprise tier.

6. Bland AI

The enterprise phone-operations play: a bundled $0.14/min talk rate with no token math, a real free on-ramp (the $0-platform-fee Start plan), and pricing that assumes serious call volume — the next tier is a $299/mo platform fee. Ranked sixth for this audience because developers prototyping will ship faster on Retell or Vapi; teams running regulated, high-volume call operations should still shortlist it.

Frequently asked questions

What is the best voice AI API for AI agents?

For most teams, Retell AI — no platform fee, component pricing totaling $0.07–$0.31/min itemized on one page, $10 signup credit. Vapi for component-level control; ElevenLabs for voice quality; LiveKit for open-source and self-hosting.

How much does a voice agent call actually cost?

Use all-in per-minute totals. Five-minute call: roughly $0.35–$1.55 on Retell (component-dependent), about $0.41–$0.65 on a default Vapi stack (per Vapi's own calculator example), $0.40 plus LLM and carrier costs on ElevenLabs once past your plan's bundled minutes ($0.08/min; LLM billed separately at unpublished per-model rates), $0.70 plus telephony on Bland. Headline fees like Vapi's $0.05/min exclude the models that make the call work.

Which voice AI API has the best free tier?

LiveKit — 1,000 agent session minutes monthly, recurring, no card. Deepgram's $200 credit is the largest one-time offer. Retell gives $10 in credits; Vapi gives $5; ElevenLabs' free plan includes 10,000 TTS credits and 15 agent minutes.

Platform or building blocks — which should I choose?

Platforms (Retell, Vapi, Bland) bundle STT + LLM + TTS + telephony: fastest to ship. Building blocks (Deepgram) are best-in-class components you assemble. LiveKit is the middle path: open-source infrastructure and an agents framework, managed or self-hosted.

Can I self-host a voice agent stack?

With LiveKit, yes — Apache 2.0 server, SIP, egress, and Agents framework; self-hosting removes platform fees and quotas. Deepgram offers self-hosted containers at enterprise tier. Vapi, Retell, and Bland are hosted only.

Sources (official pages, checked October 2026): retellai.com/pricing · vapi.ai/pricing · livekit.io/pricing and LiveKit's official blog (overage rate) · deepgram.com/pricing · bland.ai/pricing · elevenlabs.io/pricing, elevenlabs.io/pricing/agents, and elevenlabs.io/pricing/api (ElevenLabs figures checked October 14, 2026 — this entry was completed after ElevenLabs' pages were read in full; its LLM per-model rates are not published on those pages and are flagged where relevant). LiveKit's paid-tier monthly fees did not render and are marked unverified. No third-party sources were used for any figure on this page.