Real-Time AI Voice API Pricing Explained

Engineering teams building customer service voicebots, sales assistants, interactive characters, and in-app voice agents must choose between three distinct architectural paradigms: integrated multimodal end-to-end APIs (such as OpenAI Realtime API), modular streaming pipelines (combining Deepgram or Whisper for transcription, Claude or GPT for reasoning, and Cartesia or ElevenLabs for synthesis), or fully managed voice agent orchestration platforms (such as Retell AI or Vapi). Each architecture shifts cost burdens across different meters, dramatically altering monthly operational expenses at scale.

Up to five meters run at once in a live voice session: speech input, reasoning, speech output, transport and telephony

In a traditional batch TTS pipeline, pricing is straightforward: a developer sends a string of characters and pays a fixed rate per thousand or million characters. In real-time conversational voice, five distinct cost categories accumulate simultaneously during an ongoing session:

1. Inbound Audio Processing (Speech-to-Text / Audio Input)

When a user speaks into a microphone or telephone handset, their voice stream must be digitized, streamed over low-latency protocols, and converted into linguistic tokens. Providers bill this either by audio duration (for example, Deepgram Nova-3 charging $0.0077 per minute of streamed audio on pay-as-you-go) or by raw multimodal input audio tokens (for example, OpenAI's gpt-realtime billing $32.00 per million audio input tokens, which translates to roughly $0.06 per minute of user speech).

2. Reasoning and Linguistic Generation (LLM Core)

Once transcribed, the conversation context is processed by an LLM to generate the assistant's reply. In modular architectures, developers pay standard text token rates (such as $2.00 per million input tokens and $10.00 per million output tokens for Claude Sonnet 5, or $0.10 / $0.50 for GPT-6 Luna). In native multimodal realtime models, audio tokens consume reasoning capacity directly, bypassing textual transcription but incurring higher base token unit costs.

3. Outbound Speech Synthesis (Text-to-Speech Streaming)

The generated response must be converted to audio with sub-second time-to-first-byte (TTFB) latency. Providers meter outbound voice through distinct units:

• Cartesia Sonic: Charges by output audio duration or model credits, translating to approximately $0.015 to $0.020 per minute of synthesized speech with 90ms TTFB.
• ElevenLabs Conversational AI & Flash v2.5: Billed per character generated, where Flash v2.5 provides a 50% credit discount compared to flagship Multilingual v2, averaging $0.020 to $0.035 per speech minute.
• Deepgram Aura: Billed per thousand characters ($0.015 per 1,000 characters), averaging roughly $0.012 per minute of synthesized speech.

4. Transport, Connection State, and Concurrency

Maintaining continuous, bi-directional audio streaming requires WebSocket or WebRTC connections. Managed platforms frequently charge session-runtime fees (such as $0.05 to $0.08 per minute) to handle turn-taking arbitration, ambient noise suppression, endpoint detection, and speech interruption handling (barge-in). Furthermore, enterprise contracts frequently levy concurrency charges for reserved channels during peak call surges.

5. Telephony Ingestion and Termination

If voice agents operate over traditional public switched telephone networks (PSTN) rather than in-app WebRTC, telephony carrier surcharges apply. In the US, Twilio's Elastic SIP Trunking charges $0.0034 per minute for inbound local calls and $0.0100 for outbound calls to the 48 contiguous states, and Telnyx starts at $0.0032 inbound and $0.005 outbound, before toll-free surcharges and phone number leasing.

OpenAI Realtime, a self-built modular pipeline or Vapi and Retell AI, compared on latency, fees and upkeep

Understanding how vendor pricing models translate into operational budgets requires analyzing the three primary architectural designs deployed in production:

Approach A: Integrated Native Multimodal Realtime API

OpenAI Realtime API represents the integrated approach. Audio is streamed directly into the model without an intermediate textual transcription layer, and the model synthesizes conversational audio responses natively. This eliminates transcription serialization latency, resulting in fluid human-like turn-taking. However, audio tokens are expensive: $32 per million input audio tokens and $64 per million output audio tokens on gpt-realtime. While text tokens cost $4 input / $16 output per million, streaming continuous background audio quickly consumes high-cost audio tokens unless aggressive Voice Activity Detection (VAD) suppresses silence.

Approach B: Modular Low-Latency Streaming Pipeline

In a modular pipeline, developers compose best-in-breed specialized microservices: Deepgram Nova-3 for streaming STT, a small fast LLM (such as GPT-6 Luna, Claude Haiku 4.5, or a fine-tuned open-source model) for reasoning, and Cartesia Sonic or ElevenLabs Flash v2.5 for streaming TTS. The developer manages orchestration and WebSocket transport via self-hosted infrastructure (LiveKit, Asterisk, or Node.js). This approach yields the lowest raw infrastructure unit cost per minute, but requires significant internal engineering to handle interruption state machines, jitter buffers, and telephony synchronization.

Approach C: Managed Voice Agent Platform (Vapi / Retell AI)

Managed platforms provide turnkey agent infrastructure. Developers configure prompt instructions, voice IDs, and tool-calling webhooks, while the platform coordinates STT, LLM, TTS, telephony, and interruption logic. These platforms bill an orchestration fee ($0.05 to $0.08 per call minute) on top of underlying vendor costs (or allow developers to bring their own API keys). While the per-minute cost is highest, development velocity is accelerated and infrastructure maintenance is offloaded.

End-to-End Cost per Minute Simulation

To establish realistic budgetary benchmarks, we simulate a standard customer service phone call lasting exactly 5 minutes (300 seconds). In this typical call profile:

• The human customer speaks for 2.0 minutes (40% of duration).
• The AI assistant speaks for 2.0 minutes (40% of duration, generating approximately 300 words / 1,500 characters).
• Silence, pauses, and turn-taking latency account for 1.0 minute (20% of duration).
• Telephony operates over inbound SIP trunking ($0.0034/min, Twilio's US local inbound rate).

Modular Pipeline Cost Breakdown (5-Minute Call):

• STT (Deepgram Nova-3 streaming, listening across 5 minutes): 5 min × $0.0077 = $0.0385
• LLM Reasoning (GPT-6 Luna, 6 turns, ~4,000 input tokens + 600 output tokens): ~$0.0007
• TTS (Cartesia Sonic, 2.0 minutes synthesized audio): 2 min × $0.0150 = $0.0300
• Telephony Inbound (Twilio SIP): 5 min × $0.0034 = $0.0170
• Self-Hosted Compute & WebSocket Egress: ~$0.0050
• Total Call Cost: $0.0912 (Effective rate: $0.0182 per minute).

Managed Platform Breakdown (5-Minute Call via Vapi/Retell):

• Platform Orchestration Fee: 5 min × $0.0500 = $0.2500
• Underlying STT + LLM + TTS costs: ~$0.0692
• Telephony Inbound: 5 min × $0.0034 = $0.0170
• Total Call Cost: $0.3362 (Effective rate: $0.0672 per minute).

OpenAI Realtime API Breakdown (5-Minute Call):

• Input Audio Tokens (including ambient background noise, ~25,000 audio tokens at $32 per million): ~$0.8000
• Output Audio Tokens (~15,000 generated audio tokens at $64 per million): ~$0.9600
• Context Window & Text Tokens: ~$0.0150
• Third-Party WebRTC-to-SIP Bridge & Telephony: 5 min × $0.0100 = $0.0500
• Total Call Cost: $1.8250 (Effective rate: $0.3650 per minute).

OpenAI's newer GPT-Live-1 voice model offers a different meter for integrated voice: $0.05 per minute of voice session, billed per second, plus the backend model and tool usage it delegates to.

At 100,000 calls per month, the modular pipeline saves over $24,000 monthly compared to managed platforms, providing clear financial justification for engineering teams to invest in self-hosted LiveKit and custom pipeline orchestration.

Production Optimization Strategies

Organizations deploying real-time voice APIs can implement four engineering practices to significantly reduce unit costs without harming caller experience:

1. Client-Side Voice Activity Detection (VAD): Never stream continuous unmuted audio to raw audio-token endpoints. Implementing aggressive local VAD cuts inbound audio transmission during customer listening pauses, eliminating up to 50% of unnecessary audio token billing.
2. Streaming Text Buffering: In modular pipelines, do not wait for the LLM to complete its entire response before initiating TTS synthesis. Buffer the first 5 to 8 words and immediately initiate WebSocket audio streaming. This cuts TTFB below 500ms while keeping model generation on low-cost text models.
3. Model Tier Switching by Interaction Stage: Use lightweight fast models (like GPT-6 Luna or Claude Haiku 4.5) for initial greeting and intent classification, escalating to heavier reasoning models only when complex transactional logic is required.
4. SIP Trunk Direct Peering: Route telephone calls directly through private SIP interconnects rather than public carrier aggregators, lowering telephony ingestion expenses from $0.015 to under $0.004 per minute.