Voice AI & Full-Duplex Agent Cost Calculator

Simulate 1-30 min multi-turn open-mic costs (OpenAI gpt-live-1 vs gpt-realtime-2.1, Gemini 3.8 Live, Vapi/Retell, PersonaPlex-7B) plus zero-upload audio duration ASR & TTS billing.

10 min
True Native Full-Duplex Flat Tier
$0.07/min
OpenAI gpt-live-1 · $0.72/call · $1,450.00/mo
Turn-Based VAD Barge-in Tax
$0.23/min3.2× gpt-live-1
gpt-realtime-2.1 · 38 turns · 73% cache
Open-Weight Duplex Self-Host
$0.0011/min
PersonaPlex-7B INT4 (RTX 4090)VRAM →
Cost Inversion Alert: Under active open-mic interruptions, turn-based gpt-realtime-2.1 hits $0.23/min vs flat-rate native full-duplex gpt-live-1 at $0.07/min (Gemini 3.8 Live rises to $0.15/min).Hear Full-Duplex Latency Difference →

1-30 Minute Cumulative Call Cost Curve ($ USD / Session)

Highlights turn-rescan compounding vs flat wall-clock & self-hosted GPUs
Voice AI Stack / APIDuplex ArchitectureEff. $/min$ / 10m CallMonthly TotalTools
Self-Host Whisper-v3-Turbo + CosyVoice3 (RTX 4090)
Pulsed batching shares GPU slots during silence — 4× higher concurrency than full-duplex
Self-Hosted Open-Weight$0.0003/m$0.0027$5.36VRAM
Self-Host PersonaPlex-7B (INT4 on RTX 4090)
True 12.5Hz full-duplex on 1× 24GB GPU (hard-capped at 4 BF16 / 10 INT4 persistent streams; zero marginal token cost)
Tier 1 · Native Full-Duplex$0.0011/m$0.01$21.43VRAM
LiveKit Cloud + Groq Whisper + Self-Host TTS
Semantic EOU agent SDK + ultra-cheap ASR + self-hosted CosyVoice3 / Kokoro
Tier 3 · VAD Barge-in Half-Duplex$0.02/m$0.21$422.44VRAM
OpenAI gpt-live-1
$0.05/min flat wall-clock audio frontend + delegated text tokens; immune to open-mic silence & barge-in waste
Tier 1 · Native Full-Duplex$0.07/m$0.72$1,450.00API
OpenAI gpt-realtime-2.1-mini
Compact realtime S2S tier; still incurs turn-rescan & unplayed barge-in waste
Tier 3 · VAD Barge-in Half-Duplex$0.08/m$0.75$1,505.60API
Retell + Deepgram + Cartesia Sonic 2
Managed telephony + 90ms Cartesia streaming TTS + lightweight LLM
Tier 3 · VAD Barge-in Half-Duplex$0.10/m$0.98$1,963.56API
Vapi + Deepgram Nova-3 + ElevenLabs Flash
Turnkey WebRTC voice orchestration + ElevenLabs Flash TTS + GPT-6.1 Terra
Tier 3 · VAD Barge-in Half-Duplex$0.11/m$1.07$2,148.67API
Google Gemini 3.8 Live
25 tok/s (1,500 tok/min) continuous open-mic stream + uncached turn-by-turn context rescanning
Tier 2 · Hybrid Quasi-Duplex$0.15/m$1.50$2,996.55API
OpenAI gpt-realtime-2.1
Turn-based server_vad: high interruption splits turns + item.truncate breaks prefix cache ($64/1M out)
Tier 3 · VAD Barge-in Half-Duplex$0.23/m$2.33$4,659.20API

Voice AI & Full-Duplex Billing FAQ

Why does OpenAI gpt-realtime-2.1 cost more than native full-duplex gpt-live-1?
gpt-realtime-2.1 uses server_vad turn segmentation with $64/1M audio output and turn-based history rescanning. Under frequent interruptions, truncated turns degrade prefix caching while unplayed pre-generated audio is billed 100%, unlike gpt-live-1's flat $0.05/min frontend.
Do any uploaded audio files leave my browser in the ASR Batch Estimator?
No. Files are inspected locally via browser Object URLs and HTMLAudioElement metadata with zero bytes transmitted to any server.
Why does PersonaPlex-7B support fewer concurrent streams than Whisper or CosyVoice?
PersonaPlex-7B steps synchronously at 12.5Hz (every 80ms) for the entire duration of every call, holding a 3,000-step FP16 ring KV cache (~1.46GB/stream) even during silence.
When should teams switch from Vapi/Retell/ElevenLabs to self-hosted Voice AI?
Once monthly production call volume exceeds 20,000-40,000 minutes, running LiveKit + Whisper-v3-Turbo + CosyVoice3 or PersonaPlex-7B INT4 on rented GPUs cuts marginal voice costs by 70%-90%.