DEV Community

StackOps Labs
StackOps Labs

Posted on

OpenAI Realtime API vs Deepgram vs Cartesia vs ElevenLabs in 2026: The Real Per-Minute Margin Math on 100,000 Voice AI Minutes

In 2026, the standard B2B SaaS price point for AI phone and voice agents—whether for dental reception, outbound sales qualification, or logistics dispatch—has converged around $0.12 to $0.18 per connected minute (typically $0.15/minute).

Many engineering teams build a quick prototype using a flagship speech-to-speech endpoint, price their SaaS tier at $0.15/minute, and assume they will earn a 70% software margin.

Then their first enterprise client runs 100,000 call minutes in a month (1,667 hours)—and the startup discovers its gross margin is negative (-6.7%).

Let us break down the exact 2026 per-minute unit economics across Deepgram (Modular vs Bundled Voice Agent API), Cartesia Line, OpenAI Realtime API (Mini vs Flagship), and ElevenLabs Conversational AI on a 100,000-minute monthly workload.

1. Why Flagship Speech-to-Speech Destroys $0.15/Minute SaaS Margins

When you wire a telephony stream directly into a flagship native speech-to-speech model without strict session and context controls, three cost layers compound simultaneously:

  1. Audio Input & Output Token Burn: Flagship audio tokens ($32/1M input audio tokens, $64/1M output audio tokens on gpt-realtime-2.1) accumulate on every turn—and unless you aggressively truncate or cache conversation history, previous audio turns are re-processed as context.
  2. Session Wrapper & WebRTC/SIP Runtime Fees: Managed realtime session layers (+$0.05/min GPT-Live session wrapper) and SIP/PSTN telephony termination ($0.010 to $0.014/min) add fixed per-minute floors regardless of how little the user speaks.
  3. Raw Character TTS Traps: Teams that accidentally run conversational voice agents on standard character-metered TTS plans instead of dedicated per-minute conversational tiers end up paying $0.17 to $0.20/minute on voice synthesis alone.

2. The 2026 Voice AI Unit-Economics Simulation (100,000 Minutes @ $0.15/min = $15,000 Revenue)

Let us run the exact monthly P&L calculation for a Voice AI SaaS billing 100,000 monthly minutes at $0.15/minute ($15,000.00 Gross Revenue):

Voice AI Stack (2026) STT + LLM + TTS / Runtime Breakdown Total Cost / Min Monthly Infra Cost (100k Min) Gross Profit & Margin (@ $0.15/min)
1. Deepgram Modular Stack Nova-3 STT ($0.0048) + Aura-2 TTS ($0.024) + Flash LLM ($0.004) $0.0328/min $3,280.00/mo +$11,720.00 (78.1% Margin)
2. Deepgram Voice Agent API All-in bundled orchestration ($0.065/min) $0.0650/min $6,500.00/mo +$8,500.00 (56.7% Margin)
3. Cartesia Line Agent $0.060/min runtime + $0.014/min telephony + $0.004 LLM $0.0780/min $7,800.00/mo +$7,200.00 (48.0% Margin)
4. OpenAI Realtime Mini gpt-realtime-2.1-mini ($10/$20 per 1M audio tok) + $0.05/min session $0.0850/min $8,500.00/mo +$6,500.00 (43.3% Margin)
5. ElevenLabs Conversational AI $0.080/min agent tier + telephony ($0.014) + LLM ($0.011) $0.1050/min $10,500.00/mo +$4,500.00 (30.0% Margin)
6. OpenAI Realtime Flagship gpt-realtime-2.1 ($32/$64 per 1M audio tok) + $0.05/min session $0.1600/min $16,000.00/mo -$1,000.00 (-6.7% Margin)

3. Three Architectural Rules for Profitable Voice AI SaaS

  1. Enforce Edge VAD (Voice Activity Detection) and Prompt Caching: Never stream dead air into an audio-token-metered endpoint. Gate audio frames with Silero VAD at the edge and cache static system prompts so you only pay cached-token rates on multi-turn calls.
  2. Use a Modular Pipeline for Standard Intake & FAQ Turns: Route 80% of deterministic call flows (appointment scheduling, order status, intake verification) through a modular Deepgram Nova-3 STT -> Gemini Flash / Groq LLM -> Deepgram Aura-2 or Cartesia Sonic TTS pipeline at $0.0328/minute, and reserve native speech-to-speech models for high-ticket enterprise tiers billed at $0.25+/minute.
  3. Never Use Raw Character-Credit TTS Plans for Live Calls: Always provision ElevenLabs or Cartesia via their dedicated per-minute Conversational Agent tiers ($0.06 to $0.08/min) rather than raw character bundles.

Download the 2026 Voice AI Unit-Economics Dataset & Gross Margin Simulator CLI

To model your own monthly call volume, average call duration, turn count, and per-minute pricing across every major provider, grab the 2026 Voice AI Agent & Real-Time Audio Pipeline Unit-Economics Dataset (CSV + JSON) + Node.js Cost Simulator CLI:

Top comments (0)