AI · 10 min read
Vapi vs Retell vs Bland (2026): Pricing, Latency and When to Build Your Own
Vapi, Retell and Bland compared on per-minute price, concurrency, telephony, HIPAA add-ons and lock-in, with worked costs at 2,000 and 20,000 minutes.
Zain Khalid MalikCTO & Co-founder, Innovation Insight·Published 
Short answer
Vapi charges a $0.05 per minute platform fee plus each model at cost, about $0.08 to $0.13 per minute. Retell is $0.07 to $0.31 all-in depending on voice and model. Bland is a flat $0.14 per minute on Start or $0.12 plus $299 a month on Build. All three suit a first agent; above 20,000 minutes a month, or where data control matters, a self-hosted Twilio, Deepgram and LLM stack at $0.05 to $0.10 per minute starts to pay.
How each platform is priced
All three join telephony, speech-to-text, a language model and text-to-speech into a phone agent; they differ in how they charge. Vapi charges a platform fee and passes every component through at cost, so your bill depends on the models you choose. Retell charges per component with a published all-in range. Bland bundles everything except telephony into one rate and sells capacity through plans. Prices below were checked on each vendor's pricing page in October 2026 and will change; confirm before budgeting.
| Platform | Per minute (checked October 2026) | Plans and concurrency | Add-ons and extras | Pricing page |
|---|---|---|---|---|
| Vapi | $0.05 platform fee plus STT, LLM, TTS and telephony at cost. Vapi's own calculator shows Deepgram about $0.01, OpenAI $0.008 to $0.045, ElevenLabs $0.015 to $0.024, Twilio $0.008 in and $0.014 out, so 1,000 minutes comes to about $82 to $129 | Pay-as-you-go with 4 concurrent calls; Core $29 a month for 10; Pro from $999 a month for 30; Enterprise custom | Extra concurrency $10 per line a month; HIPAA add-on $2,000 a month | vapi.ai/pricing |
| Retell AI | $0.07 to $0.31 all-in: voice engine $0.055, telephony $0.015 (free with your own SIP), TTS $0.015 for Retell, Cartesia or OpenAI voices up to $0.10 for ElevenLabs v3, LLM from $0.016 (Gemini 3.0 Flash) to $0.08 (GPT-5.4); speech-to-speech $0.07 (GPT Realtime mini) to $0.38 (GPT Realtime 2.1) | 20 concurrent calls free, then $8 per line a month; $10 of free credit to start | Knowledge base, denoising, guardrails and PII removal at $0.005 to $0.01 a minute each; phone numbers $2 a month | retellai.com/pricing |
| Bland AI | Start $0.14 including LLM, STT and TTS; Build $0.12 plus $299 a month; telephony at Twilio pass-through | Start: 10 concurrent calls and a 100 calls a day cap; Build: 50 concurrent; Enterprise custom | Transfers $0.05 or $0.04 a minute | bland.ai/pricing |
| ElevenLabs Agents | $0.08 within your plan's included minutes, $0.16 beyond | Creator $22, Pro $99, Scale $299, Business $990 a month | TTS about $0.02 a minute at the current promotional rate | elevenlabs.io/pricing/api |
| Self-hosted (Twilio + Deepgram + LLM + TTS) | Roughly $0.05 to $0.10 all-in: Twilio $0.0085 in or $0.014 out, Deepgram Nova-3 $0.0048 (promotional; $0.0077 regular), a mini-class LLM under $0.02, Aura-2 TTS $0.030 per 1,000 characters | No platform fee; concurrency set by your server and provider limits; Twilio numbers $1.15 a month | You build and run the orchestration server | twilio.com, deepgram.com |
Worked monthly cost at 2,000 and 20,000 minutes
The table applies the published rates to 2,000 minutes a month (about 500 four-minute calls, a small clinic or agency) and 20,000 minutes (a busy support or outbound line). The self-hosted row excludes building and running the server.
| Platform | Assumed rate | 2,000 minutes | 20,000 minutes | Notes |
|---|---|---|---|---|
| Vapi | $0.082 to $0.129 (their calculator) | $164 to $258 | $1,640 to $2,580, plus Core at $29 for 10 concurrent calls | Cheaper with a smaller LLM and TTS; dearer with ElevenLabs and a large model |
| Retell AI | $0.109 for a typical mix (engine $0.055, telephony $0.015, TTS $0.015, GPT-5.4 mini $0.024) | $218 (range $140 to $620) | $2,180 (range $1,400 to $6,200) | 20 concurrent calls free; add-ons add $0.005 to $0.01 each |
| Bland AI | Start $0.14; Build $0.12 plus $299 | $280 on Start, plus Twilio telephony of about $17 to $28 | $2,699 on Build ($2,400 plus $299) plus about $170 to $280 telephony; Start would be $2,800 and hit the daily call cap | Flat rate makes budgeting simple |
| ElevenLabs Agents | $0.08 within plan, $0.16 beyond | $160 to $320 plus plan fee | $1,600 to $3,200 plus plan fee | Check which plan's included minutes cover your volume |
| Self-hosted | $0.05 to $0.10 | $100 to $200 | $1,000 to $2,000 | Add engineering time to build and an on-call to run it |
Latency and voice quality: what to test
Vendor latency figures will not match what your callers hear: the number depends on the models, the telephony route, where your tools are hosted and platform load, which is why we publish no benchmarks. Test each platform the same way before you commit:
- Measure turn latency on real phone calls, not the browser demo: from the caller finishing a sentence to the first audible word of the reply. Aim for under about 800 ms; above a second, callers talk over the agent.
- Test with tool calls in the loop. Looking up an order adds your API's response time to every turn; a platform that lets the agent say a filler phrase while the tool runs matters more than raw model speed.
- Interrupt it. Speak while the agent is mid-sentence and check that it stops within a few hundred milliseconds and does not finish the old reply afterwards.
- Run the same 30 scripted calls on each platform with the same prompt and model, including background noise, a long reference number and a caller who changes their mind. Score them on a rubric, not on impressions.
Telephony and bring-your-own
All three provision numbers and all three let you bring your own. Vapi passes Twilio through at cost and supports SIP trunks. Retell charges $0.015 a minute for its telephony and waives it when you connect your own SIP trunk, which is the simplest path for teams that already run a contact centre or a Twilio account. Bland bills telephony at Twilio pass-through on top of its per-minute rate. Bringing your own carrier keeps numbers, routing, recordings and consent records in your account, and means a later platform change does not mean changing phone numbers.
Tooling and integrations
- Vapi exposes the most knobs: STT, LLM and TTS provider per assistant, tools as webhooks or server URLs, and events for every stage of the call. It suits engineering teams that want to tune the pipeline and track each component's cost.
- Retell pairs a visual flow builder with a prompt-based agent, and sells a knowledge base, guardrails, denoising and PII removal as per-minute options. It suits support and booking lines built by product and engineering together.
- Bland leans towards outbound: campaigns, pathways for scripted flows and a simple per-minute rate that is easy to budget.
- ElevenLabs Agents wins on voice quality and cloning, the natural choice when the brand voice is part of the product.
- All support webhooks and function calling to your API, warm transfers, recordings and transcripts. Check the details that bite in production: keypad digit capture, handoff with a summary, several languages on one number and recording retention.
Compliance: HIPAA, recording consent and PCI
If the agent will hear protected health information you need a Business Associate Agreement from the platform and from every model and telephony provider behind it. Vapi sells a HIPAA add-on for $2,000 a month (checked October 2026). Retell and Bland offer BAAs on enterprise terms rather than a published add-on price. On a self-hosted stack you sign BAAs directly with Twilio, Deepgram and your model provider and control retention yourself. Whichever you choose: announce recording where consent laws require it, hold recorded consent for outbound calls under the FCC's robocall rules, and capture card numbers by keypad to a tokenisation service with recording paused, never into the transcript. Our HIPAA and PCI DSS pages cover the detail.
Lock-in and portability
Prompts, evaluation sets, phone numbers and tool endpoints are portable; a platform's flow builder, knowledge base and analytics are not. Keep prompts and test scripts in your repository, bring your own telephony, host tools as plain HTTP endpoints, and export recordings and transcripts on a schedule. Done that way, moving between platforms, or to your own stack, is a few weeks of work rather than a rebuild. The platforms hardest to leave are the ones whose visual editor holds your logic.
When to build your own on Twilio, Deepgram, an LLM and TTS
A self-hosted agent is a WebSocket server that takes audio from Twilio Media Streams or a SIP gateway, streams it to Deepgram Nova-3, sends transcripts to a language model with your tools, and streams the reply through a TTS provider back to the call; you write the interruption, endpointing and recording logic. It costs about $0.05 to $0.10 a minute to run and a few weeks of senior engineering to build well. Build your own when:
- Volume is above roughly 20,000 minutes a month, where the platform fee alone exceeds the cost of a server.
- Data control is non-negotiable: healthcare, finance or government work with BAAs and retention terms signed with each provider and nothing stored by a middleman.
- Voice is the product, so the orchestration layer is core intellectual property rather than a commodity.
- You need behaviour the platforms do not expose: custom endpointing, on-device models, unusual telephony or very long calls.
- You already run the engineering and on-call to keep it up.
Stay on a platform while you are validating that a voice agent works for your callers, when volume is modest, or when nobody wants to own a real-time audio server at 3 a.m. Our usual advice: pilot on a managed platform, then move proven flows to your own stack once volume justifies it. We build both; see AI voice agent development for scope and prices, from $12k to $22k for a pilot, and the AI voice agent cost guide for how running and build costs add up.
Decision table
| Situation | Pick | Why |
|---|---|---|
| First agent, under 5,000 minutes a month, engineering team in charge | Vapi | Component pricing shows where the money goes; easiest to swap models |
| Support or booking line built by product and ops people together | Retell | Flow builder, knowledge base and guardrails as options; 20 free concurrent lines |
| Outbound reminders or qualification at volume with a fixed budget | Bland | Flat per-minute rate; campaign tooling |
| Brand voice is central, or you need cloning | ElevenLabs Agents | Best voice quality; plan-based minutes |
| Over 20,000 minutes a month, regulated data, or voice is the product | Self-hosted on Twilio, Deepgram, an LLM and TTS | Lowest per-minute cost and full control, at the price of owning the server |
| Not sure it will work yet | Any managed platform for a four-week pilot | Learn from real calls before choosing an architecture |
If a chat widget would serve your users as well, start with AI chatbot development and add voice later.
Sources
- AI voice agent development services
- AI voice agent cost guide
- AI agent cost calculator
- AI chatbot development
- Vapi pricing: https://vapi.ai/pricing (checked October 2026)
- Retell AI pricing: https://www.retellai.com/pricing (checked October 2026)
- Bland AI pricing: https://www.bland.ai/pricing (checked October 2026)
- ElevenLabs API pricing: https://elevenlabs.io/pricing/api (checked October 2026)
- Deepgram pricing: https://deepgram.com/pricing (checked October 2026)
- Twilio US voice pricing: https://www.twilio.com/en-us/voice/pricing/us (checked October 2026)
- FCC: stop unwanted robocalls and texts: https://www.fcc.gov/consumers/guides/stop-unwanted-robocalls-and-texts
Need a number for your project?
Send a short brief and get a written estimate.
A senior engineer replies within one business day. No sales call required.
Zain Khalid Malik
CTO & Co-founder, Innovation Insight
Zain owns architecture, engineering standards and the platform team at Innovation Insight. He sets the bar for code quality, security and the tooling every squad ships with.
LinkedInRelated questions.
Is Vapi or Retell cheaper?
At typical settings they are close: Vapi's own calculator puts 1,000 minutes at about $82 to $129, and a typical Retell mix comes to about $0.11 a minute, or $109 per 1,000 minutes, both checked October 2026. The model and voice you choose move the price more than the platform does.
How much does Bland AI cost per minute?
Bland's Start plan is $0.14 a minute including the LLM, speech-to-text and text-to-speech, with 10 concurrent calls and a 100 calls a day cap. Build is $0.12 a minute plus $299 a month for 50 concurrent calls. Telephony is billed at Twilio pass-through on top (checked October 2026).
Which voice AI platform has the lowest latency?
It depends on the models, telephony route and tools you use more than on the platform, so test each on real phone calls with your own prompt and tools. Aim for under about 800 ms from the end of the caller's sentence to the agent's first word.
Do Vapi, Retell and Bland support HIPAA?
Vapi sells a HIPAA add-on for $2,000 a month (checked October 2026). Retell and Bland offer BAAs on enterprise terms. A self-hosted stack lets you sign BAAs directly with Twilio, Deepgram and your model provider.
When should we build our own voice agent instead of using a platform?
Above roughly 20,000 minutes a month, when regulated data requires direct control of every provider, when voice is the core of your product, or when you need behaviour the platforms do not expose. Otherwise pilot on a platform first and move proven flows later.