Vapi vs Retell vs Bland: how to choose

Vapi, Retell and Bland all host AI phone agents, but they are built on different ideas. Vapi is a layer that joins a speech-to-text provider, a language model and a voice that you choose. Retell runs the call for you and lets you shape the conversation with a prompt or a flow of nodes. Bland runs its own full stack and prices it as one rate.

Pick by the trade you want between choice and simplicity, then confirm it with your own test calls. Vendor latency figures are not comparable with each other, so measure on your own line.

Key facts

• Vapi charges $0.05 a minute for hosting, with the transcriber, model and voice billed on top (Vapi pricing).
• Retell quotes $0.07 to $0.31 a minute in total (Retell pricing).
• Bland charges $0.14 a minute on Start and $0.12 on Build, which has a $299 monthly fee (Bland pricing).
• Bland says it "owns the full voice AI stack" (Bland docs).
• Retell offers a single prompt or a conversation flow of nodes.

How do they compare at a glance?

Each row is set out with its source in the sections below. The latency figures come from the Vapi home page, the Retell home page and Bland's docs.

How is each one built?

Vapi describes itself as "an orchestration layer over three modules: the transcriber, the model, and the voice". Its introduction says each can be swapped for a provider of your choosing, and that you can plug in your own server as the LLM. An assistant is one agent. Squads split the work across several specialised assistants that hand off to each other during a call.

Retell offers a single prompt for simple agents and a conversation flow for tighter control, where the task is split into nodes. With a custom LLM, Retell handles the call itself, including telephony, transcription, turn-taking and speech synthesis, and your server writes each reply over a WebSocket.

Bland's docs say it "owns the full voice AI stack: telephony, speech-to-text, language model, text-to-speech, and pathway orchestration" (Bland docs). Its home page says it runs entirely on its own models, and that self-hosted and on-premises deployments are available for the most sensitive workloads.

How much control do you get?

The three sit at different points between choice and simplicity.

Vapi gives the most choice of parts (Vapi docs). You pick and swap each provider, which suits teams that want a specific voice or model and are happy to tune the pieces. The cost is more settings to get right.

Retell sits in the middle. Its flows give you control over the path of a call, and a custom LLM gives you the reply, while Retell keeps the audio and turn-taking.

Bland trades choice for one vendor. You do not assemble parts, and one team runs the whole stack. That can mean fewer moving pieces. We did not find a way to bring your own model, so check before you commit.

How do they connect to phone lines?

All three can use numbers you already have. Vapi's docs cover importing a Twilio number and SIP trunking. Retell lets you buy a US or Canada number without a provider account, or connect your own provider through SIP trunking or SIP URI dialling. Bland's SIP docs cover connecting your own carrier, attaching numbers and porting numbers in.

Whichever you use, test on the carrier you will use live. A transfer, a keypad menu or a voicemail can behave differently on a different route.

What do their latency claims mean?

Each vendor publishes a figure, and none of the pages we read says how it was measured. Vapi's home page claims "sub-500ms latency" and lists an average latency of under 500ms (Vapi). Its own docs give a target instead, that the whole voice-to-voice flow should ideally take under 500 to 700ms (Vapi docs). Retell's home page says "~600ms latency" (Retell). Bland's docs say "sub-400ms response latency" (Bland docs).

These may not measure the same thing. One figure might stop when the model starts replying and another when the caller hears sound. What matters is the gap the caller hears, called voice-to-voice latency, measured on your own calls with your own model and voice. Our guide to turn-taking and latency explains how. We have not measured any of the three yet. Our own benchmark figures are planned.

How do they charge?

Vapi charges $0.05 a minute for hosting, with the transcriber, model and voice billed on top (Vapi pricing). Retell adds up voice infrastructure, telephony, voice and model, and quotes $0.07 to $0.31 a minute in total (Retell pricing). Bland charges one rate per plan, from $0.14 a minute on Start to $0.12 on Build with a $299 monthly fee (Bland pricing). Our pricing comparison works through a 3-minute call on each.

What testing hooks does each one have?

Vapi has evals for decisions at one point in a conversation, and simulations where an AI tester holds a whole conversation, in chat or with full audio. Our Vapi testing guide covers both, plus call logs and webhooks.

Retell's simulation testing runs as a text conversation and grades each run. For audio you use a web call or a phone call test, which also covers keypad menus, transfers and carrier audio. See our Retell testing guide.

Bland's evaluations use judges that read a conversation, and test cases that act as simulated callers. A failing required judge blocks a version from being promoted. Its testbed takes one turn from a real call, lets you edit what is sent to the model, and runs that turn again. Its post-call webhooks send call data to your server when a call ends.

How should you decide?

Narrow the field on fit, then let your own calls decide.

1. Write down what you cannot change: the carrier, any model or voice you must use, where data may sit, and your expected call volume.
2. Rule out any platform that cannot meet those on its own pages.
3. Build the same small agent on each remaining platform, with the same prompt and the same tools.
4. Run the same set of scenarios against each, over the phone, with interruptions, pauses, noise and a transfer.
5. Time the gap the caller hears on the recordings, and mark each call pass or fail against the same rules.
6. Divide each platform's bill by the minutes you used.

Run enough calls for the pass rates to mean something. Our guide on test calls shows why ten is rarely enough. RingTru can ring each agent with the same simulated callers, so the comparison is like for like. For more options, see Vapi alternatives.

Everything here was checked against the linked vendor pages on 28 October 2026. These guides name the platforms in plain text and are not endorsed by any of them.