The Real Cost of Running an AI Voice Agent in 2026

Understanding what drives AI voice pricing is essential for anyone building or scaling a production-grade agent. The AI voice agent cost in 2026 is shaped by several core components:

Core price components:

• Speech Synthesis (Text-to-Speech / TTS) – converts written text into natural-sounding speech.
• Speech Recognition (Speech-to-Text / STT / ASR) – converts spoken audio into written text using models that transcribe language in real time.
• Large Language Model (LLM) – neural networks trained on vast text data to understand and generate human-like responses.
• Voice Agent Platform – provides the infrastructure and tools to build, deploy, and manage AI-powered voice agents.
• Transport (Telephony / WebRTC) – connects audio between user and agent, either over PSTN telephony (Twilio, Telnyx) or WebRTC (Daily, LiveKit, Agora) for browser-based agents.

Cost breakdown for a production-grade AI voice agent

1. Speech Synthesis (Text-to-Speech / TTS)

What is TTS and why is it needed?

Text-to-Speech allows a voice agent to speak responses in a natural-sounding voice, which is essential for phone-based or voice interactions. Understanding the real-time vs turn-based TTS architecture is crucial for optimizing both costs and performance.

What are the key TTS considerations?

• Voice quality and naturalness: Modern neural TTS voices sound more human-like, with attention to tone and prosody. Quality varies by provider and voice type.
• Language and voice options: Support for multiple languages and accents is crucial for global use. Custom voices (branded voice personas) may be offered at higher cost.
• Latency: The time to synthesize speech should be low to keep conversations flowing. Streaming TTS can output audio in real-time as text is processed.
• Integration: TTS is usually accessed via cloud API. Some providers allow on-premise or edge deployment (often at enterprise tiers) for low latency or privacy.

What are the usual TTS billing models?

• Per character (or per million characters) of text converted to speech. This is the most common model – you pay based on the length of the text input.
• Free tiers are common (e.g. millions of chars per month free) to get started. Beyond that, it is pure usage-based billing.

What are the usual TTS price ranges?

Roughly $4 to $200 per 1 million characters in 2026. Standard cloud voices (Polly Standard, Google Standard) sit at ~$4/M. Mainstream neural voices (Polly Neural, Google Neural2, Azure Neural, OpenAI gpt-4o-mini-tts) cluster around $12–$22/M. Premium real-time models (Cartesia Sonic 3 ~$35/M effective, ElevenLabs Turbo/Flash v2.5 ~$50/M) and ultra-realistic flagships (ElevenLabs v3 ~$100/M, Hume Octave ~$50–$150/M, Google Studio $160/M) dominate the high end. Open-source self-hosted (Kokoro, CosyVoice, Qwen3-TTS) has no per-character rate at all: the bill is the box it runs on. Translates to roughly $0.005–$0.04 per minute of generated speech.

What are the key TTS providers?

• Cartesia (Sonic 3, Sonic Turbo, Sonic 2): High-speed neural TTS optimized for conversational use. Sonic 3 is the 2026 flagship with ~90ms latency and 15 credits/sec audio billing.
• ElevenLabs (v3, Turbo v2.5, Flash v2.5): Ultra-realistic voice cloning and ~75ms streaming. v3 ~$100/M chars, Turbo/Flash ~$50/M.
• OpenAI (gpt-4o-mini-tts, tts-1, tts-1-hd): Token-based audio output ($12/M effective for mini-tts, $15/M for tts-1, $30/M HD).
• Hume Octave: Emotion-aware TTS with fine-grained prosody control. Pro tier $50–$100/M chars.
• Inworld (TTS-1.5 Mini, TTS-1.5 Max): Real-time TTS with Mini at $5/M chars and Max at $10/M, roughly $0.005-$0.01 per minute of speech.
• Rime (Mist, Arcana): Per-audio-minute billing ($0.030/min PAYG, ~$39/M chars effective). Strong for English voice agents.
• MiniMax Hailuo Speech (2.5 Turbo, 2.6 HD): Multilingual TTS ~$40–$50/M chars. Credit-based subscription.
• Microsoft Azure Speech (Neural, Neural HD): SSML support and fine-grained control. Neural $15/M, Neural HD $22/M.
• Amazon Polly: Standard $4/M, Neural $16/M, Generative $30/M, Long-form $100/M.
• Google Cloud TTS (Standard, Neural2, Chirp 3 HD, Studio): $4/$16/$30/$160 per 1M chars respectively.
• Open source self-host: Fish Speech, Kokoro, Qwen-TTS – free model weights, only GPU compute.

2. Speech Recognition (Speech-to-Text / STT / ASR)

What is STT and why is it needed?

Speech-to-Text (STT), also known as Automatic Speech Recognition (ASR), is the component that understands what the user is saying. It transcribes the caller's speech into text so that it can be processed by an LLM. Without accurate STT, the voice agent cannot know the user's request. For an in-depth comparison of STT and TTS providers, check our comprehensive STT/TTS selection guide.

Key considerations?

• Accuracy: Quality varies by model (general vs. phone-call optimized models, etc.). Good models achieve 5–10% WER, meaning the accuracy is 90–95%.
• Latency: For a live voice agent, low latency streaming transcription is needed so you are not keeping the user waiting. Represented by the Real-Time Factor (RTF) metric.
• Language support: Important if your service needs to handle multiple languages. Some STT engines specialize or have better accuracy in certain languages.
• Customization: Some providers allow custom vocabularies or even custom acoustic models (training on specific data) to improve accuracy on proper nouns or industry jargon.
• Features: Extras like speaker diarization, punctuation capitalization, profanity filtering, etc., can add value in voice agent use-cases.

Usual billing models:

• Per minute of input audio – typically billed per minute (or per second) of audio processed. Most cloud STT services charge based on the length of the audio input.
• Often pay-as-you-go, with automatic volume tier discounts. Some have free monthly minutes.
• Streaming vs batch: Some providers price streaming (real-time) and batch (asynchronous file transcription) similarly, others may differ slightly.

Usual price ranges:

Streaming STT spans roughly $0.0015–$0.024 per minute in 2026. The cheapest end is self-hosted or specialized models (NVIDIA Parakeet via Together at $0.0015/min, Cartesia Ink-Whisper at $0.00217/min). Mainstream cloud providers cluster around $0.0025–$0.012/min. Realtime/multimodal models like OpenAI gpt-realtime bundle STT+LLM at ~$0.06/min. Volume commits can drop streaming rates by 30–50%.

Key providers

• Deepgram (Nova-3): Real-time STT engine optimized for low latency and high accuracy. Standard streaming $0.0077/min mono, $0.0092/min multilingual. A $0.0048 promotional rate runs on the same page and will lapse; budget on the standard one.
• AssemblyAI (Universal-Streaming, Universal-3 Pro): API-first STT with prompt-steerable models, speaker diarization, and summarization. Universal-Streaming $0.0025/min, Universal-3 Pro Streaming $0.0075/min.
• Cartesia Ink-Whisper: Cheapest real-time streaming STT for voice agents – $0.00217/min on Scale plan.
• NVIDIA Parakeet TDT 0.6B v3: Open-weights ASR backed by NVIDIA. Available via Together AI at $0.0015/min, or self-hosted on GPU.
• OpenAI gpt-4o-transcribe / Whisper V3: $0.006/min effective. gpt-4o-mini-transcribe at $0.003/min. Whisper open weights for self-host.
• Speechmatics (Ursa 2): Strong multilingual support and accent robustness. Standard PAYG ~$0.012/min, Pro tier from $0.004/min on commit.
• Gladia (Solaria): Low-latency transcription with speaker detection. Starter $0.0125/min, Growth as low as $0.00417/min.
• Google Cloud Speech-to-Text (Chirp 2/3): Enterprise STT with wide language support, streaming and batch modes (~$0.024/min streaming).

3. Large Language Models (LLMs)

What is it and why is it needed?

The LLM is essentially the "brain" of the voice agent. It takes the transcribed text input and processes language – understanding intent and context – then generates a text response. In a modern AI voice agent, an LLM (like GPT-style models) enables natural conversation, dynamic responses, and handling of free-form user input that rule-based systems cannot. For detailed guidance on selecting the right model, see our comprehensive LLM selection guide.

Key considerations:

• Model capability: Different LLMs have different strengths and weaknesses. Choosing an appropriate model impacts how well your agent can handle complex queries.
• Context length: Newer models in 2026 support very long inputs (sometimes millions of tokens), but using long contexts can be costly.
• Latency: Large models can be slow. For voice agents, you may favor a slightly smaller or distilled model if it significantly improves response time.
• Privacy and hosting: Sending conversation data to an external API might raise compliance issues if the content is sensitive. For businesses handling sensitive data, review our guides on SOC 2 compliance and legal compliance requirements.
• Customization: Some providers allow fine-tuning (which may incur training and hosting costs).

Usual billing models:

Token-based billing – this is standard for most LLM APIs. You pay for input tokens (the text you send in, including conversation history) and output tokens (the text the model generates). A token is roughly 0.75 words, so 1,000 tokens ~ 750 words.

Key providers:

• OpenAI: GPT-5.5, GPT-5.4 family (mini/nano), GPT-5, plus gpt-realtime for bundled audio-in/audio-out. 2026 lineup spans $0.20/M input (5.4 nano) to $30/M output (5.5 Pro).
• Anthropic: Claude Opus 4.7, Sonnet 4.6, Haiku 4.5. 1M-token context window, prompt caching with 90% discount on cache hits.
• Google: Gemini 3.1 Pro, Gemini 3 Flash, Gemini 3.1 Flash Lite. Ultra-fast and cheap for voice agents – Gemini 3 Flash $0.50/M input is a common balanced choice.
• DeepSeek: V4 Pro and V4 Flash. V4 Flash $0.14/$0.28 per 1M is among the cheapest competent models.
• xAI: Grok 4.3 and Grok 4.1 Fast. Grok 4.1 Fast at $0.20/$0.50 punches above its weight for voice.
• Meta: Llama 4 Maverick / Scout, Llama 3.3 70B. Open weights – host on Together AI, Fireworks, Groq, or self-hosted GPU.
• Mistral: Mistral Large 3 ($0.50/$1.50), Medium 3.5, Small 4. Open-weights European alternative.

4. Voice Agent Platform

What is it and why is it needed?

A Voice Agent Platform is an orchestration layer or framework for building the actual voice bot/agent. It typically ties together the telephony, STT, LLM, and TTS components, handling the call flow logic, state management, and integration with any backend systems.

Key considerations:

• Integration vs. all-in-one: Some platforms require you to bring your own STT, LLM, etc., while others provide an end-to-end solution.
• Features: Look for call recording, DTMF for IVRs, real-time handoff, latency management, and analytics.
• Scalability and reliability: Handling potentially many concurrent calls is non-trivial. Platforms take care of scaling the voice infrastructure.
• Quality assurance: Implementing proper testing and monitoring is essential for production deployments. Learn more about QA metrics and testing tools for voice agents.

Usual billing models:

Usage-based (per minute of call) – The platform might charge you per minute of voice call handled by the AI agent. This often encompasses the underlying costs (STT, TTS, etc.), essentially as a bundled rate.

Usual price ranges:

$0.01–$0.14 per minute in 2026. Pure orchestration platforms (Pipecat Cloud, LiveKit Cloud Agents) charge $0.01/min and pass model costs through at vendor cost. Mid-market bring-your-own-key platforms (Vapi $0.05, Retell $0.055, Synthflow $0.09) layer an explicit per-minute fee on raw component costs. Bundled platforms (Bland AI $0.11–$0.14) embed LLM/STT/TTS/telephony in a single rate, with an estimated 15–40% markup on the underlying components. Speech-to-speech bundles (Ultravox $0.05) include the model itself.

Key providers:

• Vapi: $0.05/min orchestration fee. Bring-your-own-key for STT/LLM/TTS, developer-oriented middleware.
• Bland AI: $0.11–$0.14/min bundled (STT+LLM+TTS+telephony included). Plus $0.04–$0.05/min on warm transfers.
• Retell AI: $0.055/min voice infra, components priced separately. All-in PAYG range $0.07–$0.31/min.
• Synthflow: $0.09/min voice engine. All-in $0.15–$0.24/min. Add-ons for performance routing, edge latency, white-label.
• Millis AI: $0.02/min base + pass-through components. BYO LLM is free.
• LiveKit Cloud Agents: $0.01/min agent session. Build tier free up to 1K min/mo. Open-source self-host option available.
• Pipecat Cloud (Daily): $0.01/min active for agent-1x ($0.0005/min reserved). Built-in SIP $0.005/min, PSTN $0.018/min, transfer $0.20/event. Open-source Pipecat framework available for self-host.
• Ultravox: Speech-to-speech bundled at $0.05/min (Whisper + GLM + TTS integrated). SIP variant $0.005/min.
• Agora Conversational AI: $0.0265/min audio + ASR participant-min. First 300 min/mo free.

5. Transport (Telephony / WebRTC)

What is it and why is it needed?

Transport carries audio between user and agent. Two modes: PSTN telephony (phone calls over SIP trunks) and WebRTC (browser- or app-embedded calls). PSTN providers handle phone numbers, inbound/outbound call routing, and SIP session control. WebRTC providers handle media servers, NAT traversal, and adaptive bitrate. Pick telephony if users dial a number; pick WebRTC if they hit a button on your web or mobile app.

Usual billing models:

• Telephony per-minute – billed per call duration, rates differ for inbound vs outbound and by destination. US local typically $0.0035–$0.014/min depending on direction and provider.
• Phone number rental – monthly fee per DID. US local DIDs: $0.50/mo (Plivo, SignalWire), ~$1/mo (Telnyx), $1.15/mo (Twilio).
• Transfer / dial fees – Telnyx charges $0.10 per Dial verb invocation (warm transfers). Pipecat Cloud SIP Refer transfers $0.20 per event. Twilio has no separate transfer fee.
• Regulatory pass-through – US carriers add 5–15% in USF, state telecom, E911 surcharges on top of advertised per-minute rates.
• WebRTC per participant-minute – typically $0.0004–$0.004/min for audio-only; minimum-monthly tiers on Cloud plans (LiveKit Ship $50/mo, Scale $500/mo). Self-hosted LiveKit/mediasoup is free OSS plus your own TURN bandwidth.

Usual price ranges:

Telephony US: $0.005–$0.02/min for inbound + outbound average, plus $0.50–$2/mo per number, plus regulatory surcharge. WebRTC audio-only: $0.0004–$0.004/min on managed tiers, $0 on self-host (excluding infra).

Key telephony providers:

• Twilio: Market leader. Local in $0.0085, out $0.014. Global reach, no per-transfer fee, $1.15/mo DID.
• Telnyx: Cheaper baseline ($0.005/min outbound, in from $0.0035) but $0.10 per warm transfer can dominate cost at high transfer rates.
• Plivo: Local in $0.0055, out $0.0115; SIP/Browser-SDK flat $0.0033. $0.50/mo DID.
• SignalWire: Local in $0.0066, out $0.008; SIP/WebRTC flat $0.003. $0.50/mo DID.
• Bandwidth, Vonage, Zadarma: Other CPaaS options with regional or specialized pricing. Vonage offers per-second billing.

Key WebRTC providers:

• Daily.co: Audio-only $0.00099/min standard, drops to $0.00036/min at volume. 10K free participant-min/mo.
• LiveKit Cloud: $0.0005/min on Ship tier, $0.0004/min on Scale. Open-source self-host option.
• Agora: ~$0.001/min audio. 10K free min/mo.
• 100ms: Audio ~$0.001/min (75% off video).
• Self-host (LiveKit / mediasoup): Free OSS; cost is TURN bandwidth and compute (~$0.0001–0.0005/participant-min depending on egress).

Summary: AI voice agent cost per minute

Note: LLMs are priced per token, TTS per character, STT per minute, transport per minute or participant-minute. Per-minute estimates above assume typical voice usage (150 WPM, ~4 turns/min) and apply a 1.8× reality factor on LLM cost to capture conversation-history token growth (compounds O(n²) with turns), function-calling round-trips, and barge-in handling. Bundled platforms like Bland AI embed an estimated 15–40% markup on the underlying STT/TTS/LLM. Telephony figures exclude US 5–15% USF / regulatory pass-through and per-DID monthly rental.

All Model Pricing

A self-hosted row has no vendor rate. It requires hosting.

Large Language Models (LLM)

Text-to-Speech (TTS)

Speech-to-Text (STT)

Developer Platforms

Transport (Telephony / WebRTC)