I build voice agents that handle 50 million calls a year. I don’t build them on Vapi, Retell, ElevenLabs or Bland. I use LiveKit or Pipecat.
1. You can only do what the SDK allows
Vapi, Retell, ElevenLabs and Bland are closed-source platforms. You talk to your agent through the SDK they give you, and that’s it.
Those SDKs cover 95-99% of use cases. If you build voice agents long enough, you’ll eventually need something in the remaining 1-5%. What that is changes from platform to platform, but it always happens.
It doesn’t happen often. When it does, it’s painful. I usually found a workaround, and the workaround was always worse than what an open framework would have let me build. Those are the moments you wish you’d gone open source from the start.
2. The per-minute price is high
The LLM costs about the same wherever you run it, and so does the phone line, so I compare everything before them. The speech parts are trickier. Some platforms bundle speech-to-text and the voice into their per-minute price and some don’t, which makes them hard to compare. Where they aren’t included, I’ve added Deepgram Nova-3 for speech-to-text (about 0.8¢ a minute) and Fish Audio’s latest model, s2.1-pro, for the voice. Fish charges $15 per million bytes of text. At around 150 words a minute that’s about 1.3¢ for every minute the agent speaks, so it’s on the high side for a whole call.
| Option | Per minute with speech-to-text and voice, before the LLM and phone line |
|---|---|
| LiveKit Cloud | about 3.1¢ (1¢ agent session plus 0.8¢ STT plus 1.3¢ voice) |
| Pipecat Cloud | about 3.1¢ (1¢ hosting plus 0.8¢ STT plus 1.3¢ voice) |
| Vapi | about 7.1¢ (5¢ platform plus 0.8¢ STT plus 1.3¢ voice) |
| Retell | about 7¢ (5.5¢ infrastructure plus its own 1.5¢ voice) |
| ElevenLabs | 8¢ on every plan. Calls over your concurrency limit cost double, 16¢. |
| Bland | 11-14¢ since the December 2025 price increase. This one includes the LLM. |
Self-hosting LiveKit or Pipecat brings the first two down further. They all do the same job, so I don’t see what paying more than double a minute buys.
HIPAA adds to the bill. Vapi charges about $2K a month for its HIPAA add-on. Retell’s BAA is free to self-sign, and its PII removal is an optional extra 1¢ a minute. LiveKit signs one on its $500-a-month Scale plan. I covered what HIPAA and GDPR actually require in HIPAA and GDPR compliance for voice AI.
3. What you get in return is simplicity, and it doesn’t last
The pitch is simple: click a button and the agent is deployed. You also get built-in integrations and hosting, and everything happening under the hood is hidden from you.
Then you hit point 1, and the simple platform has turned into a fight. The framework code is hidden too, so Claude Code and other coding agents can’t help you as much as they could with an open framework. They can’t read what they’re working against.
If you’re non-technical, don’t use Claude Code, or don’t want to put much time into voice AI, these platforms can still get you an okay-ish agent. You could probably get somewhere.
4. Reliability needs testing, and testing needs access
At scale, everything along the way fails. The call doesn’t connect. The call connects, but there’s no audio. The accent switches from English to Hindi for no reason. The LLM starts thinking while it’s speaking.
The only way to stay ahead of that is to test automatically and to call your agent yourself. A closed platform makes the automated part harder. So you get the fastest route to an agent that isn’t really usable.
I hear the same story again and again. Someone gets a decent agent on Vapi, then struggles to push it further. Adding features makes it worse, regressions show up, and development slows to a painful pace. The problems compound: LLMs don’t know how the framework works, you don’t have knobs to steer it, and you can’t see enough to test it properly.
I’d rather keep moving and know that every change is covered by automated tests, then call the agent myself to catch anything a human would notice that a test wouldn’t. I wrote more about that in 5 costly mistakes I made building voice AI.
5. You can’t self-host, and you can’t leave
Self-hosting doesn’t matter to everyone. You do it to save on hosting, to keep tighter control of data privacy, and to get lower latency and better performance when it’s done right. It takes development effort.
With Vapi, Retell, ElevenLabs and Bland, it isn’t an option at all. There’s nothing to self-host. If you later change your mind about your stack, you rip out the whole agent and start from scratch on another platform. Rebuilding a full system is expensive.
What I use instead
I use LiveKit or Pipecat. Dograh is a third option. I just haven’t used it.
LiveKit is my pick when I want to self-host, because it gives me all three parts: the SIP package, the server that handles sessions, and the agent framework.
Pipecat is a really nicely engineered framework, and in my opinion it gives you even finer control over turn-taking. You can do whatever you want with it. The catch is that it’s only the framework. You bring your own SIP and server, and the docs mostly tell you to use something. That’s how they make money, and I respect it, but it means more work for you.
Stuck on a closed platform?
Hit the SDK wall, or watching the per-minute bill grow? Book a free consultation and bring the feature your platform won’t let you build. I build production voice agents on LiveKit and Pipecat. For the rest of the stack, the Voice AI Master Handbook goes out to newsletter subscribers.