Retell vs Vapi vs Bland vs Custom LiveKit (2026): Price per Minute, Latency and Lock-In

What is the difference between Retell, Vapi, Bland and a custom LiveKit stack?

Retell, Vapi and Bland are managed platforms: each bundles call transport, speech recognition, a language model and speech synthesis behind a per-minute price. A custom LiveKit stack is an architecture instead. You assemble those parts on LiveKit's open-source framework, pay each provider at list price, and own the pipeline, the on-call and the margin.

The three platforms sell different things. Retell sells the quickest route to a working phone agent, with an itemised price. Vapi sells orchestration: a platform fee, with speech, model and telephony passed through at cost, which suits developers who want to pick each part. Bland sells one flat rate that covers its own models, and sizes its plans by call volume.

I have shipped production voice on both paths. My team integrated Retell in about two days, according to the case study Retell published, and the AI voice interview step it powered cut false-positive assessments from 50% to 15%. On a custom LiveKit stack, I cut voice cost from about 10 cents a minute on a managed platform to about 2.5 cents in production. Both decisions were right at the volume we had when we made them.

How much does each option cost per minute in 2026?

On list prices checked 23 September 2026: Retell's default is $0.11 a minute before telephony, Vapi lands between about 8 and 14 cents all in, Bland charges $0.12 to $0.14 plus telephony, and LiveKit's own calculator default is about 4.8 cents. My custom LiveKit stack runs at about 2.5 cents a minute in production.

Retell's pricing page is the fairest baseline because it itemises: $0.055 a minute for voice infrastructure, a per-minute price for each model (for example $0.0064 for GPT 5.6 Luna, $0.064 for GPT 5.6 Terra or Claude 5 Sonnet, $0.16 for GPT 5.5) and $0.015 for most voices, $0.040 for ElevenLabs. Its calculator's default adds up to $0.11, and the headline range it publishes is $0.07 to $0.31.

Vapi charges $0.05 a minute for hosting and passes providers through at cost. On its own listed ranges (transcription $0.0095 to $0.0099, OpenAI models $0.0077 to $0.0452, ElevenLabs voices $0.0146 to $0.0238), the low end of those ranges on Vapi's free telephony is 8.2 cents, and the high end on Twilio outbound is 14.3 cents. Bland sells one rate covering model, transcription and voice: $0.14 on its free Start plan, $0.12 on Build at $299 a month.

LiveKit Cloud charges for the session rather than the stack: 1,000 free agent minutes on Build, then $0.01 a minute beyond the 5,000 included in Ship ($50 a month) or the 50,000 in Scale ($500 a month), with speech and model billed by the providers you choose. Its calculator's default phone agent is $0.0479 a minute. Against Retell's $0.11 default, a measured 2.5 cents is a 4.4x reduction; against LiveKit's own default, 1.9x.

What do the per-minute prices leave out?

Telephony, concurrency, add-ons and plan fees. Retell's default shows telephony at $0.00 until you add its $0.015 a minute; Bland bills telephony separately; Vapi's platform fee excludes the providers it passes through. Concurrency caps and daily call limits then decide which plan you are really on.

Read the billing rules as closely as the rate. Retell bills to the second, charges for silence and hold because transcription keeps listening, does not bill calls that fail to connect, and stops the agent fee after a transfer while telephony keeps running. Add-ons compound: a knowledge base (+$0.005), denoising (+$0.005) and PII removal (+$0.01) add $0.02 a minute, 18% on top of the default.

Plan limits are where a cheap quote turns into a plan change. Vapi includes 4 concurrent calls on usage-only billing and charges $10 a line for more; Retell includes 20 and charges $8 for each call above that; Bland caps Start at 10 concurrent calls and 100 calls a day. Size concurrency to your busiest hour, not your monthly minutes.

The line no platform prints is engineering time on the custom path. Counting two engineer-days a month of maintenance, a custom stack does not beat a vendor until about 20,000 minutes a month; the arithmetic is in voice AI build vs buy. Run your own mix of models, voices and minutes in the voice AI cost calculator.

Which is fastest, and how should you compare latency claims?

Each vendor publishes its own number: Retell says about 600ms, Vapi says under 500ms on average, Bland says 400ms. None of those pages states the method, the region or the percentile, so treat them as claims until you have run the same call script through each option yourself.

All four run the same cascade. Audio arrives, the system decides the caller has finished, speech becomes text, the model starts answering and synthesis starts speaking. On default settings the largest line is usually end-of-turn detection rather than the model, because waiting for silence is how a pipeline decides the caller is done, and that wait is a setting. The stage-by-stage budget is in voice AI latency under 800ms.

Measure the tail, not the average. Put the same 50 recorded calls through each option from the same region and carrier, and log the gap between the caller's last word and the agent's first audio at the 95th percentile. A steady 700ms feels better to a caller than a 400ms median with a long tail, because the tail is where people start talking over the agent.

The platforms tune turn-taking for you, and Retell advertises its own turn-taking model. On LiveKit you own it: the Apache-2.0 agents framework ships a transformer-based turn detector, and endpointing can be tuned per use case. That matters when your callers are unusual (older callers, noisy lines, several languages). For a first agent, a platform default is usually good enough.

How does telephony work on each, and can you keep your own numbers?

All four accept a SIP trunk from your own carrier, and all four also sell telephony: Retell at $0.015 a minute, Vapi on its own free telephony or Twilio, Bland at pass-through cost, LiveKit at $1 a month per number. Buy numbers on your own carrier account from day one, and switching platforms later stays a configuration change.

The rates: Retell charges $0.015 a minute on its numbers, $2 a month per number and $10 for a verified number. Vapi's own telephony and SIP are free, with Twilio at $0.008 a minute inbound and $0.014 outbound, and Telnyx at $0.0055. LiveKit includes one free US number, then $1 a month, and charges $0.004 a minute for third-party SIP on Ship after 5,000 included minutes ($0.003 on Scale). A direct Telnyx trunk is $0.0032 a minute inbound and $0.005 outbound on local US numbers.

The real trade is ownership. A number bought inside a platform has to be ported out when you leave, which is a carrier process with its own lead time. A number on your own Twilio or Telnyx account can point at Retell today and at your LiveKit trunk next year without a caller noticing. Trunk choice and SIP setup are in LiveKit SIP trunking: Twilio vs Telnyx.

Outbound adds rules no platform removes. The FCC's February 2024 ruling treats AI-generated voices as artificial voices under the TCPA, so calls need prior express consent, written consent for telemarketing, and must identify the business. Bland's daily caps (100 calls on Start, 2,000 on Build) are the other outbound limit to check before you plan a campaign.

How much can you customise each one, and where is the lock-in?

The prompt is portable everywhere. The rest is not: call flows, turn-taking behaviour, analytics, evaluation history and platform-bought numbers stay with the vendor. Retell and Vapi accept your own language model, Vapi lets you choose each provider, Bland runs its own models, and a LiveKit stack is code in your own repository.

Retell's custom LLM mode opens a WebSocket to your server for each call, while Retell keeps telephony, transcription, turn-taking and synthesis. For most products that is enough control: your logic, its voice pipeline. Vapi goes further, with providers you pick and a custom model endpoint. Bland's flat rate is the opposite trade: its own models, no provider pass-through, and custom voices on Enterprise.

Lock-in is the cost nobody models. Over two years, being able to swap a speech or voice provider the week a better one ships is worth more than the per-minute difference. On LiveKit that swap is a code change behind a flag; on a platform it waits for the vendor's integration list.

On a platform, three habits keep leaving cheap: numbers on your own carrier account, transcripts and recordings exported to your own storage by webhook from day one, and prompts and tool definitions kept in your repository rather than only in a dashboard. Then a move is a migration, not a rewrite, and the playbook is in migrating from Retell or Vapi to LiveKit.

Which is compliant: HIPAA, SOC 2 and data retention?

All four can support regulated work, at very different prices. Retell offers a BAA you sign yourself at no extra fee; Vapi sells HIPAA as a $2,000 a month add-on; Bland puts its BAA on Enterprise; LiveKit Cloud includes SOC 2 Type II and a BAA from its $500 Scale plan. Self-hosting makes compliance your own work.

The details matter more than the badges. Retell's compliance page lists SOC 2 Type 1 and 2 and per-agent retention from one day to two years. Vapi's HIPAA mode changes how logs, recordings and transcripts are stored, needs the paid add-on and a signed BAA, and cannot run alongside zero data retention. Bland lists SOC 2 Type I and II, PCI DSS and GDPR, with on-premises or VPC deployment on Enterprise.

For a small healthcare pilot the arithmetic is short: Vapi's add-on costs as much as 16,000 minutes on Retell at $0.125. For data that must stay inside your network, the choice narrows to Bland Enterprise on-premises or a self-hosted LiveKit stack, where the BAA chain runs through every provider you call, and a provider that stores ePHI is a business associate even when it cannot view the data.

Which should you choose, by use case and volume?

Choose by monthly minutes first, then by constraints. Under about 20,000 minutes a month a managed platform is the correct commercial answer: Retell to ship fastest, Vapi for provider control, Bland for a flat rate at volume. Past that, with engineers who can own a pipeline, a custom LiveKit stack wins on cost and control.

The arithmetic at scale is blunt. At 100,000 minutes a month, Retell's default plus its telephony is $12,500 on pay-as-you-go pricing; the same minutes on a custom stack at 2.5 cents are $2,500. That is $10,000 a month, $120,000 a year, before any enterprise discount a platform would offer at that volume, which you should ask for before deciding.

In between, run both. Put a routing layer in front, send a small share of real calls to your own stack, compare transcripts, latency and transfer rates on identical call types, and move traffic in steps. What decides whether callers notice is turn detection and barge-in, not the model. To price the product around the calls, use the AI product cost estimator.

How do you move from a platform to LiveKit without dropping calls?

Not with a cutover. Instrument the platform you have, rebuild to parity, put a routing layer in front of both stacks and shift traffic in steps, with a rollback trigger at each one. Keep the platform account live for about 60 days after the new stack carries all of the traffic.

Most of the parity work is platform features you forgot you depended on: post-call analysis, campaign dialling, knowledge base, denoising, PII redaction, branded caller ID and warm transfer. Each is a line on the platform invoice and a small project on your own stack, so list them from the invoice before you estimate the move.

If you would rather have the LiveKit stack built and instrumented for you, voice AI development is built at $0: the work is split into checkpoints with acceptance criteria agreed before it starts, and each one is invoiced only after you have seen it and accepted it.