I priced every voice-AI stack for a 10k-minute month. The results surprised me
I was working through pricing for a voice-agent SaaS and got annoyed by how hard it is to answer one basic question.
If I run 10,000 minutes of calls in a month, what do I actually pay?
Not the headline platform fee. Not the nice landing-page number. The real stack cost after you add the phone line, speech-to-text, text-to-speech, and the voice agent provider.
So I put Vapi, Retell, Bland, Twilio Voice, and OpenAI Realtime into the same model and ran a simple 10k-minute month.
The short version is that the cheapest-looking option is not always cheapest. Once you add STT and TTS, the gap moves around fast.
Here is the shape of the comparison I used.
| Stack | What is included | What you still need to check |
|---|---|---|
| Vapi | orchestration layer for voice agents | model, voice, telephony and usage assumptions |
| Retell | voice agent platform pricing | included minutes and overage rules |
| Bland | voice automation platform pricing | campaign style, telephony, and concurrency limits |
| Twilio Voice | raw telephony building block | STT, TTS, LLM, orchestration, storage |
| OpenAI Realtime | direct realtime model path | phone carrier, call control, app logic |
The mistake I see is comparing Vapi or Retell directly against Twilio Voice. That is not the same product. Twilio is mostly the carrier and programmable voice layer. If you use it raw, you still need the agent brain, streaming, speech recognition, speech generation, logs, retry logic, billing, and all the boring glue.
That glue has a cost even if the invoice does not show up as one line item.
For a 10k-minute month, the calculator makes you separate four questions.
- How many minutes are inbound vs outbound?
- Which provider charges per minute on top of model cost?
- Is speech-to-text bundled or separate?
- Is text-to-speech bundled or separate?
The surprising bit was how often the raw telephony path only wins if you already have the engineering time to own the whole stack. If you need to ship a product this week, the platform markup can be cheaper than spending two weeks debugging call streams and usage reconciliation.
The other surprise was that voice choice matters more than people expect. A nice neural voice can turn a cheap call path into a much more expensive one. Same with transcription. If you pick a premium STT provider and a premium TTS provider, your per-minute economics change before the agent provider has even charged you.
Here is a simplified example.
| Cost bucket | Why it matters |
|---|---|
| Telephony | every minute needs a phone network path |
| Voice agent platform | orchestration, latency handling, tools, call state |
| STT | turns caller audio into text |
| LLM or realtime model | reasoning and response generation |
| TTS | turns responses back into audio |
| Storage and logs | recordings, transcripts, analytics, compliance |
The pricing pages make each bucket look small in isolation. The month total is where it gets real.
This matters if you are selling voice agents to customers. A client does not care that Twilio was cheap if the full product loses money at 2,000 minutes. They also do not care that a platform is elegant if their expected call volume makes the margin disappear.
The sane way to price this is backwards. Start with the customer plan, estimate expected call minutes, add a safety margin for spikes, then choose the stack. Do not choose the stack first and hope the unit economics work later.
For example, if you sell a $199 per month voice-agent plan and the customer uses 10,000 minutes, your gross margin can get wrecked. If you sell a $999 plan with usage caps and overages, the same provider might be fine. The product packaging matters as much as the infrastructure provider.
I also think more voice-AI demos should publish their assumed monthly minutes. A demo that handles five test calls tells you almost nothing about cost at real usage. The calculator forced me to make those assumptions explicit, which made the provider comparison much cleaner.
None of this means one provider is always better. Vapi can make sense when you want a developer-first orchestration layer. Retell can make sense if its bundled path matches your use case. Bland can make sense for certain outbound workflows. Twilio plus your own stack can make sense when you have the engineering team and need control. OpenAI Realtime can make sense when you want to build close to the model.
The point is that you should compare full stack cost, not logo pricing.
I put this into a calculator so anyone can plug their own numbers: https://callforge.dev/calculator
Top comments (0)