Voice AI pricing is opaque on purpose. Vendors quote one rate, hide the per-component spend, and rely on customers not knowing what a minute of audio actually costs to run. Let's pull the curtain back.
What a minute of voice actually costs
- STT (Deepgram Nova-3): $0.0077/min wholesale
- LLM (GPT-4o mini at ~6k tokens/min): $0.0014/min wholesale
- TTS (Cartesia Sonic-3 Pro): ~$0.04/min effective
- Telephony (Twilio US): $0.01/min
- Total raw COGS: ~$0.06/min
Add infra (LiveKit, monitoring), a thin platform fee, and you land near $0.09/min for the cheapest premium combo. That's the actual floor in 2026.
Where competitors hide margin
Vapi advertises 'at cost' passthrough but adds a $0.05/min platform fee on top. Retell starts at $0.13/min with their cheapest voice, which is a $0.07 platform-fee markup over similar COGS. Floo's Starter is $0.09 with a tighter ~$0.03 fee margin.
Watch out for per-seat fees. Some platforms charge $50-200/seat/month on top of per-minute rates. For a team of 10 reps that's $500-2,000/month before a single call.
When custom combos save money
If you have your own OpenAI API key, almost every platform lets you 'bring your own' and skip the markup. On Floo this turns a $0.09 minute into a $0.072 minute (~20% savings). Worth doing once you're past 50k min/month.
