S2S vs Pipeline — voice agent cost calculator
Model your monthly voice agent bill by session length, volume, and architecture. Prices verified April 2026.
How the numbers are calculated
S2S per-minute cost rises with session length because audio tokens from all prior turns are re-billed on every turn — this is documented in Google's Vertex Live API spec and confirmed by OpenAI developer community reports. We model this with a quadratic-with-cap curve calibrated against published token rates and independent benchmarks: gpt-realtime runs roughly $0.24/min at 2 minutes, $1.00/min at 15 minutes, and reaches context window limits around $1.50/min for 30-minute sessions. Sources: OpenAI pricing page, Google AI pricing page.
Pipeline per-minute cost is flat because streaming ASR and TTS are billed per audio minute, and only compact text context accumulates for the LLM. Tiers: Economy ~$0.05/min, Standard ~$0.10/min, Premium ~$0.18/min — based on Deepgram Nova-3 ($0.0077/min), GPT-4o-mini / Gemini Flash-Lite / GPT-4o text rates, and Aura-2 / ElevenLabs Flash TTS pricing.
Excludes: telephony/SIP, orchestration platform markup, network/transport costs. These apply to both architectures equally and do not change the comparison. Approximation is deliberate: this tool is for sizing, not invoicing.