Here’s the verdict before you sit through three sales demos. Vapi, Retell, and Bland AI are developer platforms, not finished voice agents. Vapi starts at $0.05/min for the orchestration layer plus whatever your model and telephony cost, and it hands you the most control. Retell runs $0.07 to $0.31/min with infrastructure and speech bundled and your pick of LLM. Bland charges $0.11 to $0.14/min with the model, speech-to-text, and text-to-speech folded into one flat rate. Pick one and you’ve bought a toolkit, not an agent. You still design, build, integrate, prompt-engineer, test, and operate the thing yourself.
That last part is the whole point of this post. If you’ve got a capable in-house dev team, build directly on one of these and skip the middleman. We mean that. But if you don’t, gmware builds and runs the entire agent for you on top of exactly these platforms, or your own stack. Reach out and we’ll tell you which fork you’re on in one call.
We’re gmware, a software development firm headquartered in Austin, TX with engineering centers in Bangalore and Mohali, India. We build and operate AI voice and agent systems for companies that want the outcome without staffing the build. Below: what each platform actually is, the per-minute math, the decision matrix, and the honest list of everything you still own after the credit card clears.
Vapi vs Retell vs Bland at a glance
Start here, because the headline per-minute rates hide most of the real cost. Vapi’s $0.05 is a platform fee on top of models and telephony you provide. Bland’s $0.14 is nearly everything in one number. Retell splits the difference. The right column is the one buyers skip and regret.
| Vapi | Retell | Bland AI | |
|---|---|---|---|
| Pricing model | $0.05/min platform fee, models at cost | $0.07 to $0.31/min, blended by feature | $0.11 to $0.14/min, all-in |
| Model handling | Bring your own or pass-through at cost | Your choice of LLM, cost varies | Bundled, no token charges |
| Telephony | Bring your own (Twilio, etc.) | Bring your own or platform | Bundled per-minute, carrier separate |
| Concurrency | 10 lines included, $10/line extra | 20 free, $8/mo each after | 10 to 100 by plan |
| Best-fit buyer | Teams that want maximum control | Teams tuning model-per-use-case | Teams that want one predictable bill |
| What you still build | Everything below the SDK | Everything below the SDK | Everything below the SDK |
Notice the bottom row is identical. That’s not a table formatting shortcut. It’s the honest answer to “which one gets me a working agent fastest,” which is none of them, because none of them is the agent.
Per-minute rate, headline numbers
What is Vapi and who is it for
Vapi is the developer-first option. It’s an orchestration layer that sits between your telephony, your speech models, and your logic, and it assumes you want to control every piece. The Build plan is $0.05 per minute for Vapi’s hosting, and model provider costs get passed through at cost, or drop to $0 if you bring your own API key. You supply telephony (Twilio or similar). You get 10 concurrent lines included, and each additional line is $10 a month.
The compliance knobs are explicit line items. A HIPAA add-on is $2,000 a month, and zero-data-retention is $1,000 a month. The Scale plan moves to an annual contract with a fixed platform fee and committed volume, and there’s no public rate for it.
Who fits Vapi: a team that wants to swap models freely, pin its own providers, and treat the platform as plumbing. It’s the most flexible and the least hand-holding. If your engineers already know their way around STT, TTS, and LLM APIs, Vapi gets out of their way. If they don’t, that flexibility is a stack of decisions nobody on your team is qualified to make yet.
What is Retell and who is it for
Retell is the middle path. It runs a blended $0.07 to $0.31 per minute depending on the LLM and features you pick, and the components are itemized so you can see where the money goes: voice infrastructure at $0.055/min, TTS at $0.015/min for platform voices, telephony around $0.015/min on a US Twilio example, and LLM cost that varies by model. Speech-to-text and infrastructure are bundled into that per-minute rate.
Phone numbers are $2 a month. You get 20 concurrent calls free, then $8 a month per additional concurrent line. New accounts get $10 in free credits to test with, which is enough to run a few hundred test-call minutes before you commit.
Who fits Retell: a team that wants to tune the model per use case (a cheap model for FAQ deflection, a stronger one for qualification) but doesn’t want to assemble the speech pipeline from scratch. You still design the flow, write the prompts, and wire your systems. Retell just gives you a cleaner starting board than raw Vapi with sensible defaults you can override.
Not sure whether a voice agent even beats a chatbot for your case? We laid out the build-versus-buy math for text agents in our AI chatbot development cost breakdown, and the same logic mostly carries. Want us to run that math on your call volume instead of guessing? Reach out.
What is Bland AI and who is it for
Bland is the flat-rate option. It bundles LLM, speech-to-text, and text-to-speech into a single per-minute number with, in its own words, “no token charges, no model-provider pass-throughs.” That makes the bill predictable, which matters more than most technical buyers admit until finance asks why the invoice jumped.
The tiers: Start is $0.14/min with a $0 platform fee and 10 concurrent lines. Build is $0.12/min with a $299 monthly platform fee and 50 concurrent. Scale is $0.11/min with a $499 monthly fee and 100 concurrent. Enterprise is custom. Telephony still comes from your carrier arrangement.
Who fits Bland: a team that values one predictable number over model flexibility, and that’s running enough volume to earn the lower per-minute rate against the platform fee. The tradeoff is control. You take Bland’s bundled model handling rather than choosing your own, which is fine for a lot of use cases and a dealbreaker for a few. If you need a specific frontier model or on-prem inference, Bland isn’t your platform.
Bland AI plans: rate vs platform fee vs concurrency
One aside on the wider field: Synthflow is a sibling platform aimed more at no-code builders, pay-as-you-go roughly $0.11 to $0.24/min with enterprise contracts starting around $30,000/year. We compared it head-to-head against these two in a separate Synthflow vs Retell vs Vapi post if the no-code angle matters to you.
What you still build after buying the platform
This is the section vendors don’t put in their decks. Buying Vapi, Retell, or Bland gets you the orchestration layer. It does not get you an agent. Here’s the work that lands on your desk the morning after you sign up.
What the platform does NOT include
- ▢ Conversation design (what the agent says, when, and why)
- ▢ Prompt engineering plus an eval suite you re-run on every change
- ▢ CRM, calendar, and helpdesk integration (real API work)
- ▢ Telephony wiring, numbers, and call routing
- ▢ Human escalation: context transfer and queue logic
- ▢ Monitoring: latency, containment, transcript review
- ▢ Compliance handling if you carry PHI or PCI data
Take one line: prompt engineering. A voice agent that reads a wrong appointment time back to a caller doesn’t throw an error. It fails politely, at scale, and you find out from a bad review. So you need an evaluation suite that catches regressions every time you touch a prompt or swap a model, and building that suite is real engineering nobody sells you.
Then there’s integration. Booking an appointment means the agent talks to your calendar. Qualifying a lead means it writes to your CRM. Each of those is an API integration with auth, error handling, and retry logic, and it’s the same work whether you’re on Vapi, Retell, or Bland. We wrote about that integration tax in general in our guide to the cost to add AI to existing software. Voice doesn’t discount it.
And the agent has to know when to give up and hand off to a person. A caller who’s angry, a request that’s out of scope, a payment dispute: the agent recognizes the moment, transfers the call with context intact, and doesn’t dead-end anyone. That escalation flow is design work, and skipping it is how you turn a support tool into a churn engine.
When you should just build on the platform yourself
We’re a software firm and we’re about to tell you when not to hire us. Here it is. If you have engineers who are comfortable with REST APIs, prompt iteration, and a bit of telephony, build directly on Vapi, Retell, or Bland. That’s literally what these platforms are for. They exist so a competent team doesn’t have to stand up a speech pipeline from raw components. Paying anyone to sit between your team and the SDK is waste.
The DIY case is strongest when your use is contained: one clear workflow (outbound appointment reminders, inbound FAQ routing), modest volume, no gnarly compliance, and a team with the time to own it after launch. In that world the platform’s docs plus a couple of engineer-weeks gets you live, and you keep full control of tuning forever.
The build-for-me case is the mirror image. You don’t have that team, or they’re heads-down on the core product. You need it live in weeks, not quarters. The workflow touches three systems and a human handoff. You carry HIPAA or PCI obligations and don’t want to design audit logging from scratch. Or, honestly, you just want someone accountable for the thing at 2am when a carrier route flaps. That’s the fork where a firm earns its fee. If you’re comparing a voice agent to a live answering service, our AI receptionist cost guide and the answering service cost breakdown both have the numbers to check your math.
How gmware builds and operates the whole agent
When the build-for-me fork is the right one, here’s how we work. We build on whichever platform fits your case (Vapi when you need model control, Retell when you’re tuning per use case, Bland when predictable billing wins) or on your own stack if you’d rather not add a vendor. Then we own the parts the platform doesn’t: the conversation design, the prompts and their eval suite, the CRM and calendar integrations, the telephony wiring, the escalation logic, and the monitoring after launch.
We run production data systems ourselves. Shield Suite, our retail-intelligence product, tracks beverage-alcohol brands across 60,000+ storefronts, so the guardrails-and-monitoring part of this isn’t theory for us. Delivery runs through senior engineers in Bangalore and Mohali with architecture and accountability in Austin, on hours that overlap yours.
There’s no productized price for this, and we won’t pretend otherwise. A voice agent’s cost depends on your call volume, your integrations, and your compliance load, so we scope and quote it per engagement. Tell us what you’re trying to automate on the phone. Reach out and we’ll give you a straight answer on scope, cost, and timeline within 48 hours, including whether you’d be better off just building it yourself.