The Real Cost of a Voice-AI Minute: LLM, STT, TTS and Telephony

A minute of voice-AI conversation costs roughly $0.09 to $0.14 when you assemble the stack yourself. That range comes from five line items: an orchestration platform, speech-to-text (STT), a language model (LLM), text-to-speech (TTS) and telephony. Below, each line has a published rate, a date and a source, and the totals show their arithmetic.

We build PulsyAI, which charges a flat $0.10 per minute on its Starter plan. We'll compare against that number honestly, including where building it yourself comes out cheaper.

Key Takeaways

• Assembled from published October 2026 prices, one call minute costs $0.0891 (low build), $0.1028 (mid build) or $0.1403 (high build).
• TTS and the platform fee are the two largest lines in a cheap build. The LLM is the biggest swing factor: about $0.0011 to $0.0286 per minute depending on the model.
• Telephony is a small line, between $0.0052 and $0.0140 per minute for US calls in these examples, and it varies by carrier and direction.
• Vapi's own estimate for 1,000 minutes is $82 to $129, or $0.082 to $0.129 per minute (vendor-reported), which matches our range.
• A flat rate is a bet on the average. If your calls are short and your stack is cheap, you may pay more than the parts cost; if they are long, you may pay less.

What are the five parts of a voice-AI minute?

Every voice agent runs the same loop. The caller speaks, STT turns the audio into text, the LLM decides what to say, TTS turns the reply into audio, and a carrier moves the call. An orchestration layer coordinates the loop and handles interruptions. Each part is billed differently, so the "price per minute" you see on a landing page is usually a sum of components.

1. Orchestration platform: the software that runs the conversation.
2. STT: billed per minute of audio.
3. LLM: billed per token, which you must convert to minutes.
4. TTS: billed per character or per minute of generated audio.
5. Telephony: billed per minute by the carrier, plus a monthly fee per number.

How much does each part cost per minute?

Orchestration platform

Vapi lists a hosting fee of $0.05 per minute, with model costs passed through at cost (vapi.ai/pricing, October 2026). Retell lists voice infrastructure at $0.055 per minute (retellai.com/pricing, October 2026). If you self-host an open-source stack, this line is zero on the invoice and large in engineering hours.

Speech-to-text

Deepgram lists Nova-3 monolingual streaming at $0.0048 per minute and Nova-3 multilingual at $0.0058. The page shows these as lower promotional rates against earlier prices of $0.0077 and $0.0092 (deepgram.com/pricing, October 2026). Because the lower rate is promotional, the mid and high builds below use $0.0077.

Language model

LLMs are billed per million tokens, so you need an assumption for how many tokens a minute uses. The example below is illustrative, not measured:

• 4 conversational turns per minute.
• Each turn sends about 2,500 input tokens (the system prompt plus conversation history).
• Each turn returns about 60 output tokens.
• Per minute: 4 × 2,500 = 10,000 input tokens and 4 × 60 = 240 output tokens.

Prices below are per million tokens, from the vendors' own pages in October 2026 (OpenAI, Anthropic). The math is tokens ÷ 1,000,000 × price.

This ignores prompt caching, which lowers the input cost where your platform supports it, and it ignores long calls, where history grows. Platforms publish their own per-minute LLM figures too. Retell lists GPT 5 mini at $0.008 per minute on its standard tier (vendor-reported), which sits close to the $0.0086 above.

Text-to-speech

Cartesia's Sonic model bills credits at 1 per character, and its page says one minute of audio uses 750 to 800 credits (cartesia.ai/pricing, October 2026). Using 750 credits per minute:

• Pro plan: $5 ÷ 100,000 credits × 750 = $0.0375 per minute.
• Startup plan: $49 ÷ 1,250,000 × 750 = $0.0294 per minute.
• Scale plan: $299 ÷ 8,000,000 × 750 = $0.0280 per minute.

ElevenLabs lists its Flash and Turbo models at $0.04 per 1,000 characters, which its page equates to about $0.04 per minute, and its v2 Multilingual model at $0.08 (elevenlabs.io/pricing/api, October 2026). These are per minute of generated speech. If your agent talks about half the call, the real cost per call minute is lower, so the totals below are conservative.

Telephony

Twilio lists US local inbound at $0.0085 per minute, outbound to US numbers at $0.0140 and toll-free inbound at $0.0220, plus $1.15 per month per local number (twilio.com, October 2026). Telnyx lists its Voice API at $0.002 per minute and SIP trunking at $0.0032 inbound and $0.005 outbound, plus $1.00 per month per local number (telnyx.com, October 2026). Adding the Voice API and SIP inbound rates gives $0.0052. That sum is our arithmetic. The Telnyx page doesn't state it as a single rate.

If you bring your own carrier, the setup is described on our SIP trunk page and on the features page, which lists Static IP and Registrar trunk types and says "Use your own rates — no Twilio markup".

What does a full minute cost when you add it up?

Here are three example builds. Each uses published rates from above, with the LLM token assumption stated earlier. They are examples, not recommendations.

Check the low build: $0.0500 + $0.0048 + $0.0011 + $0.0280 + $0.0052 = $0.0891. In that build the platform fee is $0.05 ÷ $0.0891 = 56.1% of the cost, and TTS is $0.0280 ÷ $0.0891 = 31.4%. The LLM is 1.2%. That is why, once you are already on a small model, switching models again barely moves the total.

Three cross-checks from the platforms themselves, all vendor-reported:

• Vapi estimates $82 to $129 per month at 1,000 minutes, which is $0.082 to $0.129 per minute (Vapi pricing page, October 2026).
• Retell's worked example totals $0.11 per minute, against a headline range of $0.07 to $0.31 (Retell pricing page, October 2026).
• VoiceAIWrapper lists provider rates of $0.05 per minute for Vapi, $0.07 to $0.31 for Retell and $0.08 for ElevenLabs Agents (VoiceAIWrapper pricing page, October 2026).

Our $0.0891 to $0.1403 range overlaps all three.

How does that compare to a flat per-minute rate?

PulsyAI's Starter plan is a flat $0.10 per minute with no bring-your-own-key requirement, and the providers are included in the platform cost (features page, October 2026). The features page lists OpenAI, Claude, Grok and a fine-tuned PulsyAI model for the LLM, and Deepgram, Speechmatics and ElevenLabs for STT. It lists PulsyAI Voices, ElevenLabs, Cartesia, OpenAI and Hume for TTS.

Against the three builds above:

• Low build: $0.10 − $0.0891 = $0.0109 more than the parts cost, or 12.2% above them.
• Mid build: $0.10 − $0.1028 = −$0.0028, so the flat rate is 2.7% below the parts cost.
• High build: $0.10 − $0.1403 = −$0.0403, so the flat rate is 28.7% below the parts cost.

So a flat rate is not always cheaper. A well-tuned low build can undercut it by about a cent a minute. What you'd also be paying for is the engineering time to wire five vendors together, the white-label layer and the rebilling. Whether that is worth a cent depends on whether you sell the result to clients. Our pages don't say whether carrier minutes are included in the $0.10, so confirm that before you quote.

What should you watch for when you model your own minute?

1. Billing increments. Some providers round up each call to a minimum unit. Check the rounding before you assume a 30-second call costs half a minute.
2. Promotional rates. Deepgram's lower Nova-3 rate is labeled as limited-time on its page. Model the list price in your base case.
3. Per-number fees. $1.15 per month per Twilio number is small, but 50 numbers is $57.50 per month, which is a fixed cost to add to your margin math.
4. Concurrency. Retell lists $8.00 per extra concurrent call per month and Vapi lists $10 per additional line per month. Peak hours decide how many you need.
5. Compliance add-ons. Vapi lists HIPAA compliance at $2,000 per month on its pricing page (October 2026). A bundled platform may price this differently, so compare the same requirement on each.

How do you use this in your own pricing?

Take the total for the build you'd actually run, then add your fixed fees and your support time. Our post on pricing an AI receptionist for your agency walks through the margin math for four packaging models, using the $0.10 rate as the cost. For this post the takeaway is narrower: know your own per-minute cost to within a cent before you quote a client. Rerun the table with your own model, voice and carrier. If you resell, compare the result with the flat rate on our features page and with the other platforms you are considering.

Disclosure: PulsyAI publishes this post and sells a flat-rate product that competes with several vendors named above. All third-party prices were read from the vendors' own pages in October 2026 and can change. The token usage, talk share and build combinations are illustrative assumptions, not measurements.