Retell AI pricing: what 10,000 voice minutes cost

Retell advertises $0.07 to $0.31 a minute. Your rate depends on the voice, the model, the phone line and a handful of billing rules that change long calls more than short ones.

Duplex Compute · October 13, 2026 · Rates and versions checked October 13, 2026

How Retell builds a per-minute price

On the pay-as-you-go plan there is no platform subscription. Retell's pricing page splits a voice agent's minute into components. The default pipeline charges $0.055 per minute for Retell voice infrastructure, plus a text-to-speech voice and an LLM, each with its own per-minute price. Retell Platform Voices cost $0.015. ElevenLabs Flash costs $0.040. At the standard tier, LLMs run from $0.0016 for GPT 5 nano to $0.32 for GPT 6 Astra.

A separate Speech-to-Speech tab prices OpenAI's realtime models by the minute: GPT Realtime 2.1 at $0.38 and GPT Realtime 2.1 mini at $0.07. That tab shows the model line alone, with no voice-infrastructure row, so the examples below price it as listed. At $0.38 it also sits above the $0.31 top of Retell's headline range.

Telephony is its own line. The page lists US Twilio numbers at $0.015 per minute and charges nothing for SIP trunking or custom telephony, which means your own carrier bills those minutes instead. Retell tracks each call to the second and does not round up per call. New accounts get $10 of credit.

Three configurations at 10,000 minutes

The spread between A and C is four times. The model line drives it. Infrastructure and telephony stay at $0.07 a minute in both pipelines, so a cheaper LLM or voice is where configuration A saves.

These are list-price arithmetic, not invoices. Enterprise plans have custom pricing, and Retell's billing exceptions (below) can raise the billed duration.

Add-ons and monthly items

Retell sells call features as per-minute add-ons. Knowledge base costs $0.005 per minute, as do advanced denoising and safety guardrails. PII removal costs $0.01. AI quality assurance costs $0.10 per minute after the first 100 free minutes. Batch calls cost $0.005 per dial, and branded caller ID $0.10 per outbound call.

Adding denoising and PII removal to configuration A adds $0.015 a minute, or $150 per 10,000 minutes. That is as much as the telephony line.

Monthly items sit outside the per-minute rate. Retell phone numbers cost $2 a month each. The first 10 knowledge bases are free, then $8 a month each. With five numbers and a peak of 30 concurrent calls, configuration A at 10,000 minutes comes to $978 + $10 + $80 = $1,068.

Transfers, silence and hold

Retell's FAQ states that the AI voice agent fee stops once a call transfers, and only the telephony fee continues. Take a six-minute call where the agent talks for two minutes and then hands off to a person for four. On configuration A the AI charge covers two minutes ($0.166) and telephony covers six ($0.09), for about $0.26. Without that rule the same call would cost $0.59.

Silence and hold time are billable. The FAQ says the speech-to-text engine keeps listening for the whole call. A ten-minute call with three minutes on hold pays configuration A's AI rate for those three minutes too, about $0.25 of the call's $0.98.

Failed calls are free. A call that reaches voicemail bills only while the AI agent is active on the line.

call cost = ai_minutes × (infra + tts + llm + add_ons)
          + connected_minutes × telephony
ai_minutes stops at transfer; connected_minutes runs to hang-up

Billing exceptions

Two rules adjust the billed duration. Calls shorter than 10 seconds that use a dynamic opening message bill a 10-second minimum. Agents whose prompt context exceeds 4,000 LLM tokens have their billed duration scaled by tokens ÷ 4,000. Retell's own example bills a 60-second call with 4,800 tokens as 72 seconds.

The token count includes the global prompt, tool descriptions, node prompts, the transcript so far, tool call results and retrieved knowledge base content. Long calls and tool-heavy flows grow it as they run. A call whose context averages 6,000 tokens bills at 1.5 times its length. Calls end after one hour by default, and agent settings can raise that to two.

Concurrency

Each workspace includes 20 concurrent calls. More cost $8 per concurrent call per month. A workload that peaks at 50 simultaneous calls pays 30 × $8 = $240 a month.

Burst mode lets calls exceed the limit up to the lower of three times the limit or the limit plus 300. Each burst call carries a $0.10 per minute surcharge for its entire duration. On configuration A that more than doubles the rate of a burst call, so sustained peaks are cheaper as purchased concurrency.

When a managed platform is worth it

Retell earns its infrastructure fee when you use what it runs: phone numbers, warm and cold transfers, IVR navigation, batch campaigns, branded caller ID, post-call analysis, simulation testing and a dashboard where people who don't write code can edit an agent. Building those yourself costs engineering time and someone on call.

The fee is harder to justify when the agent lives inside your own web or mobile app. You already have the interface, the user accounts and the backend the tools call. There you are paying phone-platform rates for a voice session.

That second case is where a voice API fits. Duplex Compute is a full-duplex voice API with tool calling and long conversations at $0.01 per connected minute. It has no phone numbers, transfers or call-center dashboard, so it is not a Retell replacement for phone operations. The $0.01 is a voice-session rate, and your tool handlers and any telephony are billed separately. Vapi alternatives sorts these options by layer.

Related notes

• Voice agent pricing: calculator and billing formulas
• Vapi alternatives for developers
• LiveKit pricing for voice agents
• GPT-Live vs. Realtime API