Direct answer
Pick Vapi when engineers must own every STT/LLM/TTS layer. Pick Retell when you need turnkey telephony and dialog flows fast. Pick OpenAI Realtime when speech-to-speech quality and custom tools/MCP in-session are the product. None of them replace senior-owned evals, CRM sync, or human handoff. See OpenAI Realtime and Voice AI Systems.
Comparison table
| Factor | Vapi | Retell | OpenAI Realtime |
|---|---|---|---|
| Philosophy | Developer BYOK orchestration | Turnkey telephony + flows | Speech-to-speech model session |
| Best for | Custom stacks, provider swaps | Fast inbound/outbound phone agents | Natural barge-in, tools in live audio |
| You still own | Latency tuning, compliance design | Ops policy, CRM truth, evals | Transport, secrets, tools, evals |
| Weak fit | Non-technical launch in days | Maximum low-level audio control | Batch TTS or offline transcription only |
How we decide with clients
- If the product is a voice experience you will differentiate on, start Realtime (WebRTC or SIP).
- If the job is phone ops automation and speed-to-live matters, evaluate Retell.
- If you need provider flexibility and already have engineers for the voice stack, evaluate Vapi.
- Always add evals, warm transfer, and cost meters before scale.
Insight: Platform choice is week one. Production ownership is the engagement.
