Retell AI Pricing: Calculate One Outbound Call Before Dialing

Retell AI prices voice agents by the minute, but the useful number is not one advertised rate. The current total depends on the voice infrastructure, text-to-speech voice, language model, telephony route, optional add-ons, and billable duration selected for the call.

As of September 13, 2026, Retell advertises AI voice-agent usage from $0.07 to $0.31 per minute. For a rough upper bound, multiply $0.31 by the maximum call length. A three-minute outbound call would therefore have a $0.93 planning ceiling before any separate fixed per-call feature.

That ceiling is conservative. The exact call can cost less, but an AI agent should not dial first and explain the bill later. Inspect the current contract, preflight the exact destination and task, approve one maximum, then reconcile the provider-reported cost after the call reaches a terminal state.

Retell AI Cost at a Glance

These figures are a dated example from Retell's current pricing page, not a universal quote. The same page shows no Retell telephony charge for SIP trunking or custom telephony, but the carrier can still bill you directly. Do not compare a bundled managed-number rate with a custom-carrier rate as if they include the same services.

For the four example components above, the simple estimate is:

$0.055 voice infrastructure
+ $0.015 standard voice
+ $0.040 language model
+ $0.015 US Twilio telephony
= $0.125 per minute

A three-minute connected call at that illustrative configuration would begin at $0.375. Selecting a more expensive model or voice, another telephony route, or an add-on changes the result. Use the live page and the exact provider contract when making a purchase decision.

Calculate Cost Per Call, Not Just Cost Per Minute

The basic formula is:

estimated call cost = billable minutes x selected per-minute components
                    + fixed per-call components

For a campaign, include attempts and the actual duration distribution:

estimated campaign cost = sum of estimated call costs for every attempt

Do not multiply the average connected-call duration by the number of contacts and stop there. Your operating model should distinguish:

• attempts that never connect;

• short connected calls;

• completed calls;

• transfers or unusually long calls;

• redials and duplicate attempts;

• fixed per-dial or per-call add-ons; and

• provider credits or carrier charges that sit outside the visible voice rate.

Suppose a reviewed task allows one call with a three-minute maximum. At Retell's published range, the base planning band is:

The high-end number is a budget ceiling, not a prediction that every call will cost $0.93. It is useful because it gives the agent a hard stop before dispatch even when terminal duration and provider usage are still unknown.

Two Billing Exceptions Can Break a Naive Estimate

Retell's current billing-exceptions documentation describes two cases that deserve explicit tests.

First, a call shorter than ten seconds can still have a ten-second minimum when it uses a dynamic opening message. That makes duration x rate incomplete for the shortest calls.

Second, prompts above 4,000 language-model tokens can scale billed duration in proportion to prompt length. A large agent prompt can therefore raise the model portion even when the spoken call duration is unchanged.

These are policy inputs, not footnotes. Record whether the opening sentence is dynamic and measure the final prompt token count before estimating a large campaign. Refresh the calculation when the prompt, model, voice, telephony route, add-ons, or call cap changes.

Preflight One Outbound Call Without Dialing

Aident Loadout can inspect the current Retell outbound-call Action and validate one bounded input before dispatch. Start with account and Vault status, then discover the current public Action by the job rather than relying on a remembered name:

aident account auth status
aident vault vault --action status

aident capabilities search \
  --query "place one authorized Retell outbound task call with a strict duration limit" \
  --types '["action"]' \
  --targetEnv staging

Copy the current public Action name from discovery and inspect its schema. Then preflight the exact request:

aident capabilities preflight \
  --name "<CURRENT_PUBLIC_RETELL_OUTBOUND_CALL_ACTION>" \
  --input '{
    "toNumber":"+12025550123",
    "taskPrompt":"Ask a fictional test recipient to confirm whether Tuesday at 10:00 AM Pacific works. Do not make a purchase or collect sensitive data.",
    "callerIdentity":"Aident editorial pricing test",
    "constraints":"This is a schema-only preflight. Do not dispatch a call.",
    "desiredOutput":"Whether the input validates and what cost basis is returned.",
    "capture":"none"
  }'

The number above is in the reserved fictional 555-01xx range. It is suitable for validating the schema, not for placing a real call. Replace it only after the real recipient, purpose, authority, timing, and applicable calling rules are approved.

On September 13, that exact input passed preflight without dispatch. The quote used Retell's published $0.31-per-minute maximum and Aident's server-enforced three-minute call limit:

maximum = $0.31 per minute x 3 minutes = $0.93

The preflight returned a dynamic zero-to-$0.93 range. Settlement occurs from the terminal provider-reported combined cost, so the final charge can be lower than the cap. No call was placed during this audit.

Preflight is a cost boundary, not calling authorization. A valid schema and acceptable quote do not prove that the number may be called, the recipient consented, the task is lawful, or the agent should retry after an uncertain dispatch.

Require One Approval Record Per Call

Before an outbound call, record six things:

1. Recipient: the approved E.164 number and how it was obtained.

2. Purpose: the exact job the agent may attempt.

3. Authority: who approved the call, timing, and any required consent.

4. Limits: maximum duration, spend, retries, commitments, and sensitive-data boundaries.

5. Evidence: what may be retained after the call, with capture off unless separately approved.

6. Terminal result: call status, provider-reported cost, and the smallest evidence needed to prove the outcome.

Treat retry authority separately. If a dispatch returns an error after the provider may have accepted it, automatically dialing again can create a duplicate call. Reconcile the first call or obtain fresh authorization before another attempt.

The same separation applies to recording and transcripts. Permission to call is not permission to record. Keep capture set to none unless the user has confirmed the applicable participant consent and retention boundary.

Forecast With Real Duration Buckets

After a bounded test, replace one average with a small duration model. For example:

Apply the selected component rate and fixed add-ons to each bucket, then compare estimated and settled cost. Track at least:

• cost per attempt;

• cost per connected call;

• cost per completed task;

• duplicate-attempt rate;

• average and 95th-percentile connected duration; and

• spend by model, voice, telephony route, and add-on.

Cost per completed task is the stronger operating metric. The cheapest minute is not useful when the call fails, creates a bad handoff, or requires a human to repeat the work.

Use How to QA AI Phone Calls Beyond the Transcript to evaluate the conversation, audio, system changes, and handoff as separate evidence layers. If you are still choosing the voice platform, the Retell AI vs ElevenLabs comparison owns that broader decision. This page owns the narrower job of estimating and bounding one Retell outbound call.

Put the Bound in the Agent Policy

A safe first policy is small:

• one approved US destination;

• one exact task and caller identity;

• no purchase or sensitive-data collection;

• three-minute maximum;

• no transcript or recording capture;

• no automatic retry;

• $0.93 maximum quote; and

• terminal cost reconciliation before another call.

Expand only after the first call's outcome and cost are reviewed. Campaign-level permission should not be inferred from one successful task call.

For the difference between cost approval and consequence approval, continue with the AI agent risk-versus-spend workflow. For Aident credits, provider pricing, and BYOK behavior, see How Aident Loadout Credits, BYOK, and Action Pricing Work.

Calculate One Retell Call Before Dialing

Follow https://aident.ai/SETUP.md

Set up Aident Loadout, inspect the current Retell outbound-call Action, and preflight one fictional bounded input without executing it. Success means you can explain the selected cost layers, maximum duration, quote ceiling, authorization boundary, capture setting, and terminal settlement rule before any real number is dialed.

Set up Loadout and calculate one bounded Retell call

Sources

• Retell AI pricing, reviewed September 13, 2026.

• Retell AI billing exceptions, reviewed September 13, 2026.

• Retell outbound caller cost breakdown, reviewed September 13, 2026.

• Aident Loadout setup, reviewed September 13, 2026.

• Live Aident Loadout staging Action contract and exact no-dispatch preflight, inspected September 13, 2026.

Refresh this guide when Retell changes its published rate range, component prices, telephony routes, billing exceptions, or add-ons, or when Aident changes the outbound-call duration cap, quote, authorization, capture, retry, or terminal-settlement contract.