Retell AI vs Vapi Pricing: The Costs Hidden Behind a Per-Minute Rate

A five-cent platform fee is not a five-cent phone call

Choose Retell first when you want an integrated starting point with room for more simultaneous calls. Choose Vapi when controlling the speech, model and voice components is a requirement, and someone on your team can manage those provider relationships. Neither choice can be priced by the hosting label alone.

A builder trying to keep a voice agent below $500 per month wanted 24/7 coverage across time zones. Being available all night and talking continuously all night are different requirements. Minutes accumulate while calls run; a number being available does not by itself establish that workload.

Define monthly connected minutes and the busiest simultaneous-call period separately. Five thousand minutes spread evenly through a month is not operationally the same as five thousand minutes concentrated around a daily booking rush.

Split the budget into processing, transport and fixed costs

Total monthly cost = AI voice processing
  + phone transport during AI calls
  + phone transport after transfers
  + numbers and concurrency additions
  + support or compliance packages
  + optional analysis features

Retell's component table separates infrastructure, voice and model selection. Vapi adds provider pass-through costs to its $0.05 hosting rate. Its cost documentation explains why an estimated per-minute figure can move with input/output tokens and generated speech.

Keep provider-paid invoices in the calculation when you bring your own keys. A zero provider charge on the platform invoice can mean the provider billed you elsewhere, not that the model became free.

What 1,000 and 10,000 minutes look like before phone transport

The following figures retain each vendor's displayed configuration rather than pretending the two calculators used identical models. They are processing budgets before carrier charges and optional packages.

The ten-thousand-minute column is a tenfold usage calculation, not a volume-discount offer. For a matched comparison, select the same speech provider, language model, voice and call destination. Then run your own prompt through each platform and keep the billable components visible.

That matched prompt matters. A long knowledge-base insert or a verbose response changes token or speech usage, even if the phone conversation lasts the same number of minutes.

Long prompts can change the meaning of “per minute”

Retell documents a context-length billing exception: when prompt context exceeds 4,000 tokens, the LLM billing duration scales with context size. Its example turns a 60-second call with 4,800 tokens into 72 seconds of LLM-billable duration. The exception is about the model component; it does not establish a matching multiplier for every infrastructure or carrier charge.

Knowledge-base text, tool definitions and conversation history can all make the context larger. Loading every company policy into every call may therefore affect both clarity and spending. Retrieve the relevant material for the task rather than using maximum context as a substitute for a precise flow.

The same documentation also describes a minimum for dynamic AI-first openings on very short calls. Include short abandoned calls in a trial instead of testing only long, successful conversations. A budget built from completed bookings can miss the processing spent on people who hang up.

Transferred calls can keep creating phone charges

Retell's pricing FAQ says the AI voice-agent fee stops once the call is transferred, while telephony charges continue for the transferred portion. It also bills silence and hold time during the AI call. Four minutes of active dialogue inside a seven-minute connected call should not be budgeted as four billable minutes.

Suppose a call spends three minutes with AI and seven minutes with a staff member. Separate the AI-processing duration from the ten-minute phone connection. Apply the carrier's rates to the actual billed legs and destinations. Do not carry the AI rate across the entire transfer, or drop the remaining phone time just because a human answered.

Vapi's free US numbers are an inbound convenience, not free call usage. The documentation also limits portability and outbound use. If outbound dialing or keeping your existing number is required, include the imported number and carrier arrangement in the design from the start.

Concurrency and retention can outweigh a small minute-price difference

Vapi's package documentation makes clear that a package does not replace the usage bill. Its Core package changes concurrency and raw retention, while higher packages add different support and control arrangements. An inexpensive per-minute configuration can still be the wrong purchase if four callers fill its baseline concurrency.

For healthcare use, price the approved provider chain and the required contractual setup. Vapi's HIPAA documentation specifies eligible arrangements and a signed BAA; it also distinguishes HIPAA mode from zero-data-retention mode. Those are procurement conditions, not interchangeable checkboxes.

Judge the pilot by completed work and the full invoice

Measure a repeatable job such as appointment booking or a documented account lookup. Include an interruption, background noise, a failed tool call, a transfer and a caller who hangs up early. Count completed actions, human follow-up minutes and all chargeable components.

A builder comparing platforms around 3,000 US call minutes also asked about development and maintenance effort. That is part of the purchase: a cheaper component configuration can cost more staff time to diagnose and restore.

Retell is the stronger initial choice when baseline concurrency and an integrated operating setup matter most. Vapi is the stronger fit when provider-level control is the reason for buying a platform rather than a finished receptionist service. Decide after the same job runs reliably on the selected configuration—not after comparing two unmatched headline prices.