Every voice AI project starts with the same argument: managed platform or build it yourself. The comparison posts you find are mostly written by one of the vendors, so here is the version from someone who builds these for clients and has no platform to sell you.
The short answer
Start on Vapi or Retell. Move to LiveKit or Pipecat when one of four specific things becomes true, not before. Most teams migrate too early, spend three months rebuilding, and end up with an agent that is cheaper per minute and worse to talk to.
The four options, side by side
| Vapi | Retell | LiveKit | Pipecat | |
|---|---|---|---|---|
| Type | Managed platform | Managed platform | Framework + cloud | Open-source framework |
| Time to first agent | Hours | Hours | Days to weeks | Days to weeks |
| Typical all-in cost/min | $0.13 - $0.25 | $0.12 - $0.22 | $0.05 - $0.11 | $0.04 - $0.10 |
| Model choice | Platform list | Platform list | Anything | Anything |
| Telephony | Built in | Built in | SIP, you configure | You bring it |
| Scaling & ops | Theirs | Theirs | Yours | Yours |
| Custom logic mid-pipeline | Limited | Limited | Good | Total |
Cost figures are all-in estimates covering speech recognition, model, synthesis and telephony at moderate volume. Your real number depends heavily on which voice provider you choose. Text-to-speech is usually the single largest line item, not the language model.
Vapi
The most developer-friendly of the managed platforms. Good API, good docs, broad provider support, and you will have something answering a phone number the same afternoon. The trade is that the pipeline is not yours: when latency is 1.4 seconds and you need 800 milliseconds, there is no layer for you to go optimise.
Choose it when: you are validating a use case, you have one clear conversation flow, and volume is under roughly 20,000 minutes a month.
Retell
More opinionated than Vapi, with stronger built-in workflow and routing logic, which makes it a better fit for non-engineers on the team. Similar ceiling for the same reason: you are operating inside someone else's pipeline.
Choose it when: your conversation has real branching logic and you want that expressed in a product rather than in code.
LiveKit
A mature WebRTC platform with an agents framework on top, plus SIP telephony. This is what most teams should move to when they outgrow a managed platform. You get real-time transport that is already battle-tested at scale, and you get back control over model choice, per-minute economics and data handling.
What you inherit: worker deployment and autoscaling, turn-detection tuning, SIP trunk configuration, and observability. That is genuinely several weeks of work and it is the part migration quotes leave out.
Choose it when: volume makes per-minute cost material, you need a model or voice the platform does not offer, or you need infrastructure and data under your own control.
Pipecat
A Python framework that models a conversation as frames moving between processors. The reason to pick it over LiveKit is that you can insert your own logic anywhere in the chain: redact sensitive data before it reaches a vendor, run a moderation classifier in parallel with the response, swap the model mid-conversation, or route different intents down entirely different paths.
The reason not to pick it is that nothing is deployed, scaled or monitored for you. A Pipecat pipeline assembled from examples works beautifully on a laptop and falls over at ten concurrent calls.
Choose it when: you have a genuine requirement to intervene between pipeline stages. See Pipecat pipeline development. If you cannot name that requirement in one sentence, use LiveKit.
The four reasons to actually migrate
Notice that "we want more control" is not on that list. Control is the mechanism, not the reason. If you cannot point at one of these four, staying put is the correct engineering decision.
What I usually build
For most production systems: LiveKit for transport and telephony, with a chained speech-to-text, language model and speech synthesis pipeline inside it rather than a speech-to-speech model. The chained approach costs 200 to 400 milliseconds more, and buys a clean transcript at every turn, free model choice, and cheaper economics on long calls. When compliance asks what the agent said last Tuesday, you have an answer.
If you are weighing this up for a real product, I am happy to look at your numbers. I will tell you if the honest answer is to stay on the platform you are already paying for. If you have already decided, here is how I run a migration off Vapi or Retell.