An AI voice agent that answers calls, handles bookings, or processes inbound leads has four cost layers that most people don't fully account for until they get their first billing statement. Telephony, speech-to-text, text-to-speech, and LLM tokens all charge per minute or per call — and they stack up fast.
This post breaks down every component, gives you real per-minute costs for the main providers, and helps you figure out whether a hosted platform (Vapi, Retell, Bland) or a custom build makes sense for your volume and requirements.
The four cost layers of every AI voice agent
Before looking at specific providers, understand the stack:
- Telephony — the phone network layer (Twilio, Vonage, Telnyx). Charges per minute for inbound/outbound calls.
- Speech-to-text (STT) — transcribing what the caller says into text. Charges per second or per minute of audio.
- Text-to-speech (TTS) — converting the agent's text response back to audio. Charges per character or per minute.
- LLM inference — the model that generates the agent's responses. Charges per token.
Every call incurs all four costs simultaneously. Add them up and a 3-minute call costs roughly $0.10–$0.35 depending on your choices.
Telephony costs
Twilio: The default choice. Inbound calls are $0.0085/minute on their standard tier, outbound is similar. Phone number rental is $1–$3/month per number. Twilio is reliable and well-documented, but it's not the cheapest option.
Telnyx: ~$0.004/minute — roughly half of Twilio. Requires slightly more setup but is increasingly popular for cost-sensitive voice applications.
Vonage (now Vonage by Ericsson): Similar pricing to Twilio, slightly worse developer experience.
Bandwidth.com: Enterprise-tier pricing, usually requires volume commitments. Worth it at 50,000+ minutes/month.
For a 3-minute average call at Twilio pricing: $0.0255. At Telnyx: $0.012. Over 1,000 calls/month, that's $25.50 vs. $12 in telephony alone.
Speech-to-text costs
Deepgram Nova-3: $0.0043/minute. This is the dominant choice in 2026 for voice agents — best-in-class latency (often under 300ms), strong accuracy on phone audio, and a simple WebSocket streaming API. At 1,000 calls of 3 minutes each, Deepgram costs $12.90/month.
OpenAI Whisper (via API): $0.006/minute. Good accuracy but higher latency than Deepgram for streaming use cases — Whisper processes audio in batches, making it less ideal for real-time conversation where you want to respond quickly.
Google Speech-to-Text v2: $0.016/minute (standard). More expensive, but has strong support for multiple languages and accents.
AssemblyAI: $0.0043–$0.0064/minute depending on model. Competitive with Deepgram, includes useful features like speaker diarization.
Self-hosted Whisper: Free after hardware costs. A single A10G GPU (~$1.50/hour on Lambda Labs or CoreWeave) can transcribe hundreds of concurrent streams, making self-hosting viable at 10,000+ minutes/month.
Text-to-speech costs
TTS is where costs vary the most dramatically between "sounds robotic" and "sounds human."
ElevenLabs (Turbo v2.5): $0.0008/character (Starter plan) to $0.0003/character (Creator plan). A typical 150-word response is ~900 characters = $0.00072–$0.00027. Fast enough for real-time voice (50–90ms latency), high quality. At 1,000 calls with 10 agent turns each averaging 150 words: $7.20–$2.70/month.
OpenAI TTS-1: $15 per million characters = $0.000015/character. Much cheaper than ElevenLabs, but noticeably more synthetic. TTS-1-HD at $30/million is better but still behind ElevenLabs on naturalness.
Deepgram Aura (TTS): $0.0150/1,000 characters = $0.000015/character. Same price as OpenAI TTS-1, purpose-built for voice agents, very low latency (~50ms).
Cartesia (Sonic): $0.000065/character, ultra-low latency (30–50ms TTFB). Strong choice for latency-sensitive applications.
Eleven Labs Flash v2.5: Lowest latency product from ElevenLabs (~75ms TTFB), slightly lower quality than Turbo but the difference is minor on phone audio.
For most voice agent deployments, Deepgram Aura or Cartesia Sonic give the best cost/latency combination. ElevenLabs is worth it when voice quality is a differentiator (customer-facing brand experiences, for example).
LLM costs
Every caller turn generates tokens: you send the conversation history plus a system prompt plus the transcribed caller input, and the model generates a response.
A typical turn might have:
- System prompt: 500 tokens
- Conversation history (last 4 turns): 600 tokens
- Current caller input: 50 tokens
- Model response: 100 tokens
Total per turn: ~1,250 tokens. A 3-minute call with 5 caller turns uses ~6,250 tokens.
GPT-5 at $2.50/$10 per M tokens: ~$0.014/call Claude Haiku 4.5 at $0.80/$4 per M tokens: ~$0.005/call GPT-5 mini at $0.15/$0.60 per M tokens: ~$0.0011/call Llama 4 70B via Groq at $0.59/$0.79 per M tokens: ~$0.004/call
LLM model choice is the second-biggest lever after call volume. GPT-5 mini and Llama 70B handle most inbound call scenarios without meaningful quality loss.
Total operating cost: 1,000 calls/month
Here's a realistic breakdown for 1,000 calls averaging 3 minutes each, using mid-tier component choices (Twilio + Deepgram + Deepgram Aura + GPT-5 mini):
| Component | Cost per call | Monthly (1,000 calls) | |-----------|-------------|----------------------| | Telephony (Twilio) | $0.0255 | $25.50 | | STT (Deepgram Nova-3) | $0.0129 | $12.90 | | TTS (Deepgram Aura) | $0.002 | $2.00 | | LLM (GPT-5 mini) | $0.0011 | $1.10 | | Total | $0.041 | $41.50 |
Swap GPT-5 mini for GPT-5 and add ElevenLabs:
| Component | Cost per call | Monthly (1,000 calls) | |-----------|-------------|----------------------| | Telephony (Twilio) | $0.0255 | $25.50 | | STT (Deepgram Nova-3) | $0.0129 | $12.90 | | TTS (ElevenLabs Turbo) | $0.0072 | $7.20 | | LLM (GPT-5) | $0.014 | $14.00 | | Total | $0.059 | $59.60 |
Both are in the $40–$60/month range for 1,000 calls. At 10,000 calls/month, multiply by 10.
Platform options: Vapi, Retell, Bland
Rather than wiring up telephony + STT + TTS + LLM yourself, hosted platforms abstract the stack. Here's how they compare:
Vapi charges $0.05/minute for hosted calls (all components included) plus your LLM costs on top. For a 3-minute call: $0.15 + LLM. At 1,000 calls: $150 + LLM ($1–$14). Total: $151–$164/month. Vapi is the fastest way to get a voice agent running — their developer experience is excellent, and they support most LLMs. Not HIPAA-compliant.
Retell AI is known for the lowest latency in the hosted tier — average response latency of 800–1,000ms end-to-end vs. Vapi's 1,000–1,200ms. Pricing is similar to Vapi at ~$0.04–$0.06/minute. Retell is the right choice when conversation feel matters most. Not HIPAA-compliant by default.
Bland AI is the cheapest hosted option at around $0.09/minute flat (LLM included for their standard model). At 1,000 calls of 3 minutes: $270/month. Higher per-minute cost than Vapi or Retell but simpler pricing with no LLM overage surprises. Good for simple call flows (appointment reminders, lead qualification). Voice quality is lower than Retell.
Custom build (your own stack): $40–$60/month in operating costs for 1,000 calls as shown above. Significantly cheaper per call at scale, but requires 150–300 hours to build properly. Full control over every component.
| Platform | Cost per min | 1K calls/month | Latency | HIPAA | Custom logic | |----------|-------------|---------------|---------|-------|-------------| | Vapi | ~$0.05 + LLM | $150–$165 | ~1,100ms | No | Limited | | Retell | ~$0.04–$0.06 | $120–$180 | ~900ms | Limited | Limited | | Bland | ~$0.09 flat | $270 | ~1,200ms | No | Limited | | Custom | ~$0.04–$0.06 | $41–$60 | 700–1,000ms | Yes (configurable) | Full |
Build costs: platform vs. custom
Vapi/Retell/Bland implementation: 20–60 hours to set up a production deployment. You're configuring prompts, integrating with your CRM/calendar, testing edge cases, and deploying. At $100–$150/hour: $2,000–$9,000. But you're locked into their stack and pricing.
Custom voice agent: 150–300 hours for a production build. You're integrating telephony, building a WebSocket audio pipeline, writing the conversation manager, handling interruptions, testing latency, and deploying. At $100–$150/hour: $15,000–$45,000. But you own every component and can swap providers when pricing changes.
At WayFind Labs, custom AI voice agent development typically runs $20,000–$45,000 for a production-ready system, which pays back the cost premium over hosted platforms within 6–18 months at moderate call volumes.
When to use a platform vs. when to custom build
Use Vapi or Retell if:
- You need something live within 2–4 weeks
- Call volume is under 5,000/month
- You don't have strict data privacy requirements
- The use case is fairly standard (appointment booking, FAQ answering, lead qualification)
Build custom if:
- You need HIPAA compliance or VPC-isolated infrastructure
- Call volume will exceed 10,000/month (cost savings become significant)
- You need complex business logic — multi-step workflows, database writes, CRM updates, conditional call routing
- Voice quality is a brand differentiator and you want to control TTS choices
- You're building a platform (i.e., your product is the voice agent)
Hidden costs
Conversation design. Writing good voice agent scripts is genuinely hard. Phone conversations have no visual cues, no ability to show formatted text, and users interrupt constantly. A skilled conversation designer adds 20–40 hours to any project. Skipping this step is the most common reason voice agents get bad reviews.
Interruption handling. Real callers interrupt mid-sentence. Handling barge-in correctly (stopping TTS, re-running STT, regenerating) requires careful WebSocket pipeline design. This takes 20–40 hours to get right and is a common failure mode in quick implementations.
CRM and calendar integrations. The actual value of a voice agent comes from it updating your CRM, booking appointments, or looking up order status. Each integration is 5–15 hours. For 3–4 integrations, budget 30–60 hours.
Post-call processing. Transcription storage, call scoring, summary generation, and routing follow-up tasks all need to be built. Another 20–40 hours.
The bottom line
A hosted platform (Vapi, Retell, Bland) gets you a voice agent in 2–4 weeks for $2,000–$9,000 in build cost, with per-call operating costs of $0.12–$0.27 inclusive of LLM. It's the right starting point for most teams.
A custom build costs $15,000–$45,000 to build but drops operating costs to $0.04–$0.06 per call and gives you full control over HIPAA compliance, vendor selection, and complex logic. The crossover point where custom build pays back is roughly 8,000–15,000 calls/month over 12 months.
The number that matters most isn't the build cost — it's the cost per resolved call. A well-designed voice agent that handles 80% of inbound calls without human intervention changes the math entirely. Focus on containment rate first, cost optimization second.