We built the same AI voice agent — an inbound customer support agent with FAQ retrieval and CRM logging — on VAPI, Retell AI, and Pipecat. We measured latency, cost, customisability, and production-readiness. Here are our results.
Platform Overview
| Factor | VAPI | Retell AI | Pipecat |
|---|---|---|---|
| Type | Managed API platform | Managed API platform | Open-source framework |
| Hosting | VAPI cloud | Retell cloud | Self-hosted |
| Pricing model | Per-minute + LLM costs | Per-minute + LLM costs | Infra costs only |
| Setup time | < 30 minutes | < 30 minutes | 2–4 hours |
| Customisation | Good via hooks | Good via hooks | Unlimited (open source) |
| PSTN support | Yes (built-in SIP) | Yes (built-in SIP) | Yes (via LiveKit SIP) |
| Open source | No | No | Yes (Daily.co) |
Latency Results
We measured end-to-end latency: from the moment the user stopped speaking (VAD cutoff) to the first audio byte of the agent response. All tests used GPT-4o as the LLM and Deepgram Nova-3 as the STT.
| Platform | P50 Latency | P95 Latency | P99 Latency |
|---|---|---|---|
| VAPI | 480ms | 720ms | 950ms |
| Retell AI | 420ms | 680ms | 890ms |
| Pipecat (self-hosted, same region) | 380ms | 610ms | 820ms |
| Pipecat (LiveKit + GPT-4o Realtime) | 280ms | 450ms | 620ms |
Pipecat with GPT-4o Realtime (audio-in, audio-out) is the fastest option — but it requires self-hosting and significantly more engineering effort.
Cost Comparison (per 1,000 minutes)
| Platform | Platform fee | LLM cost (GPT-4o) | STT cost (Deepgram) | TTS cost (ElevenLabs) | Total est. |
|---|---|---|---|---|---|
| VAPI | $15 | $18 | Included | $12 | ~$45 |
| Retell AI | $12 | $18 | Included | $12 | ~$42 |
| Pipecat (self-hosted) | $8 (infra) | $18 | $4 | $12 | ~$42 |
Customisation
VAPI and Retell AI both offer webhooks for custom logic — you can intercept messages, inject context, and call your own APIs. But you cannot change the underlying media pipeline. With Pipecat, you have full control: custom VAD thresholds, custom silence detection, parallel function calling, custom audio processing, and integration with any STT/LLM/TTS provider.
Which Should You Choose?
- VAPI: Best for getting a working agent in production quickly. Excellent documentation and active community. Choose this for MVPs and proofs-of-concept.
- Retell AI: Slightly lower latency than VAPI in our tests. Better out-of-the-box support for multi-agent workflows. Good for production at moderate scale.
- Pipecat: Choose this when you need maximum customisation, minimum cost at high volume, or specific compliance requirements that prevent using managed cloud platforms.
For most teams, start with VAPI or Retell AI — the managed infrastructure means you can focus on the agent logic, not the media pipeline. Move to Pipecat when you hit cost ceilings or need customisations the managed platforms cannot support. Our AI voice team has built production agents on all three and can help you choose and implement the right stack for your requirements.