Retell AI vs Vapi (2026)

What do Retell and Vapi actually cost per minute?

Both platforms advertise a number that no production agent runs at, because both price the platform separately from the models that do the talking. Here is how the stack builds.

Retell starts at $0.055/min for voice infrastructure, adds TTS ($0.015/min for platform voices, $0.04/min for ElevenLabs), an LLM ($0.003/min for GPT-5 nano up to $0.16/min for GPT-5.5), and telephony (about $0.015/min on managed US Twilio). Retell's own pricing page frames the honest range as $0.07 to $0.31 per minute. Add-ons are itemized too: knowledge base +$0.005/min, denoising +$0.005/min, PII removal +$0.01/min. You get $10 in free credits at signup and 20 concurrent calls included.

Vapi charges a flat $0.05/min platform fee and passes speech-to-text, LLM, and text-to-speech costs through at provider cost, or at $0 to Vapi if you bring your own API keys and pay the providers directly. That makes the floor look lower, but third-party cost breakdowns put typical production deployments at roughly $0.30 per minute all-in, in the same territory as Retell. Concurrency starts at 10 calls, with extra lines at $10 per month. The compliance add-ons are the sleeper line items: HIPAA at $2,000/mo and zero-data-retention at $1,000/mo.

For a 500-minute month, expect roughly $65 to $165 on either platform depending on model and voice choices, before engineering time. If that number needs to be predictable, Retell's one-page component pricing is easier to model; if it needs to be minimal and you already have model API contracts, Vapi's pass-through wins.

Which is easier to build on?

Retell positions as a platform for building, testing, deploying, and monitoring agents: a visual Conversation Flow builder for structured call logic, single- and multi-prompt agents, a testing playground, simulation testing, and post-call analysis. A capable operations person can ship a working agent without writing code, and developers still get a full API underneath.

Vapi describes itself as the developer platform for voice agents, and the product follows: server SDKs across Node, Python, PHP, Go, and Rust, client SDKs for web and mobile, Squads for multi-assistant orchestration, and bring-your-own everything (models, keys, telephony). The dashboard can stand up a basic assistant in minutes, but tools, webhooks, and production hardening assume engineering time. Reviewers on both platforms converge on one theme: support is community-first below enterprise tiers, so your team is the first responder either way.

What about latency and call quality?

Advertised numbers are nearly identical: Retell claims about 600 ms; Vapi's pages claim under 500 ms in one place and sub-600 ms in another. Treat both as best-case laboratory figures. User reports for both platforms describe 800 to 1,000 ms typical turns with occasional spikes, and the biggest variable is your own configuration: a slow LLM adds more delay than either platform's infrastructure. Concepts like voice activity detection and barge-in matter as much as raw latency for how natural the agent feels, and both platforms expose tuning for them.

Who should pick which?

• Pick Retell if you want the shortest path to a monitored, testable production agent; if a non-engineer will maintain call flows; or if you need HIPAA with a self-serve BAA without a five-figure add-on.
• Pick Vapi if you have engineers, existing LLM API contracts, or a custom voice experience to build into an app; the SDK surface and BYO-key economics are the best in the category.
• Pick neither if what you actually need is your business phone answered. Both are construction kits; below a few thousand minutes a month, the engineering time usually costs more than the software.

If you don't want to build it yourself

Retell and Vapi sell you the factory. If you just need the car, a done-for-you AI receptionist skips the build entirely. Tilk composes the agent from your business profile, answers from your knowledge base with sub-second turn-taking, books appointments onto Google Calendar, and hands you transcripts and lead alerts, and you can talk to a live preview agent for 60 seconds before creating an account. To be clear about fit: Tilk is not a developer platform, and if you need custom voice flows inside your own software, Retell or Vapi is the right category. For the receptionist job itself, see what an AI receptionist covers and how AI answering services work.