AI voice agent platforms compared, with one I built myself as the baseline

I built an AI phone receptionist from separate parts and ran it for about five months in 2026. That gives me a real cost per minute to hold the hosted platforms against. I have not run a business on Vapi, Retell, Bland or ElevenLabs agents. For those four, this page reports their own published prices and nothing I can't see on their sites.

The baseline: four vendors and my own server

A caller dials a Twilio number. Twilio streams the audio to my server. Deepgram transcribes it. A small model on Groq writes the reply. Cartesia speaks it. The audio goes back down the same stream.

Here is one inbound minute at today's list prices:

PartVendorCost per call minute
Inbound local callTwilio$0.0085
Media StreamsTwilio$0.0044
Speech-to-text, streamingDeepgram Nova-3$0.0048
ModelGroq, gpt-oss-20babout $0.0006
Text-to-speechCartesia Sonic, $49 planabout $0.018
Totalabout $0.036

Two of those lines are estimates and you should see the assumptions. The voice line assumes the agent talks about half the time, roughly 450 characters per minute, which is the figure I budgeted with. The model line assumes four turns a minute at 1,500 tokens in and 80 out. My line ran llama-3.1-8b-instant, which Groq now prices by contact only, so the row uses gpt-oss-20b. The phone number adds $1.15 a month. The server was one I already rented.

So my running cost was under 4 cents a minute. Keep that number, because it is the misleading one.

The hosted platforms, as they publish it

PlatformPublished priceWhat is extraCalls at once, entry plan
Vapi$0.05 per minute hostingTranscriber, model, voice, phone. Each passed through at cost4, then $10 per line a month
Retell AI$0.055 per minute voice infrastructureModel, voice, phone. Their page gives $0.07 to $0.31 all in20, then $8 per call a month
Bland$0.14 per minute on Start, $0.12 on the $299 Build planPhone carrier10 on Start, 50 on Build
ElevenLabs Agents$0.08 per minuteBurst rate of $0.16 over your concurrency limitDepends on plan
Deepgram Voice Agent$0.075 per minute standardPhone carrierUp to 60 on Growth in North America
Cartesia agents$0.03 per minuteModel usage, plus $0.014 per minute on their number8 on the free plan
Twilio ConversationRelay$0.07 per minuteThe call itself, and your modelNot listed

I have run the Twilio, Deepgram and Cartesia building blocks in production. I have not run any row in this table as a bundled agent product. These are the vendors' numbers, read on November 14, 2026.

A few things in the fine print are worth more than the headline rate:

  • Vapi's $0.05 is only the platform. Their own calculator puts 1,000 minutes at $82 to $129 a month once speech and model costs are added. That is 8 to 13 cents a minute all in, on Vapi's own phone transport.
  • Retell prints a worked example: 4 cents for the model, 5.5 for infrastructure, 1.5 for voice, for 11 cents a minute on your own phone carrier.
  • Bland's Start plan caps you at 100 calls a day. That is a ceiling you hit on a good marketing day, long before cost matters.
  • Vapi lists HIPAA compliance at $2,000 a month. If you are a clinic, that line is your real price.
  • Retell scales the bill when prompts average over 4,000 tokens. A long knowledge base in the prompt raises your per-minute cost.

Why 4 cents is the misleading number

Compare my 3.6 cents with Retell's 11-cent example and building it yourself looks like a 7-cent saving on every minute. At 1,000 minutes a month, that is $70. At 5,000 minutes, $350.

Now the other column. To get to a receptionist that could hold a real call, I fixed these:

  1. Custom call parameters arriving in a different message than I expected.
  2. The agent transcribing its own greeting as caller speech and answering itself.
  3. Speech recognition settings that had changed between model versions.
  4. A one-second minimum on end-of-speech detection, which I had to mask with filler audio.
  5. Tiny text fragments turning into noise in the voice output.
  6. Safari sending a recording format that cannot be decoded live.
  7. A routing bug that only appeared when a second phone number was added.
  8. An environment file change that removed two vendor keys and muted the line.
  9. A prompt-injection hole that scripted test calls found in July.

Each one has a paragraph on the Twilio, Deepgram or Cartesia page. None was hard. Together they were spread across three months of work on that line, and a hosted platform has already paid that bill.

On my line the caller heard a filler phrase about half a second after they stopped talking and the real answer in one to four seconds. A hosted platform that does nothing but voice has had more time to shave that than I had.

When building it yourself still wins

  • Volume. At 20,000 minutes a month a 7-cent gap is $1,400 every month. The engineering is a one-time cost.
  • The call has to do something custom mid-conversation. Mine built its prompt per tenant from that business's profile and kept a per-business memory of past calls. Hosted platforms support tool calls. You are still working inside their turn loop.
  • You need to pick each vendor. A client with a data agreement that names its speech vendor cannot use a bundle that picks for them.
  • You already have the pipeline. I do. For my next phone agent the build cost is mostly paid.
  1. How much does an AI voice agent cost per minute?

    Published prices on November 14, 2026 run from about 3 cents a minute before model costs on Cartesia to 14 cents a minute on Bland's Start plan. My self-built agent cost about 3.6 cents a minute to run, not counting the engineering time to build it.

  2. Is it cheaper to build your own AI phone agent?

    Per minute, yes. In total, only at volume. The gap between my self-built stack and a hosted example was about 7 cents a minute, which is $350 a month at 5,000 minutes. Below that, the engineering time costs more than the platform fee.

  3. Which voice agent platforms has Ben Jacobs used in production?

    Twilio Media Streams, Deepgram Nova-3, Groq and Cartesia Sonic, wired together on my own server. I have not run Vapi, Retell, Bland or ElevenLabs agents in production, and their rows on this page are their own published prices.