What it is
Retell AI is a developer platform for building AI phone agents: you write the prompt or draw the call flow, pick the LLM and the voice, attach a phone number or your own SIP trunk, and pay per connected minute. It sits in the same build-it-yourself layer as Vapi, not the managed-service layer of PolyAI or Parloa — Retell hands you the speech pipeline, telephony, and testing tools, and your team owns the agent. Founded in 2023 and backed by Y Combinator (Winter 2024 batch), it is run from Palo Alto by CEO Bing Wu.
The growth is the reason it is on shortlists. Retell reported 40M+ calls a month in December 2025, and Brex named it a Top 25 fastest-growing startup in its Summer 2026 benchmark (17 August 2026). Sacra estimates annualized revenue at roughly $80M in August 2026, up from about $45M at the end of 2025 — on approximately $5.1M of disclosed funding, all of it seed money from 2024. Those revenue figures are third-party estimates, not company filings.
Why it shows up in CX and support stacks
- The cheapest way to put a real agent on a real phone line. The voice infrastructure, a choice of GPT, Claude, or Gemini models, and Twilio or Telnyx numbers are billed on one meter, with no platform fee and no annual contract on pay-as-you-go. A support team can have an after-hours line or an appointment-booking agent in production in days, not a quarter.
- Two ways to build, one per kind of call. Single-prompt agents handle open conversation; Conversation Flow is a node-based builder for structured, high-stakes calls — authentication, payments, intake — with warm transfer, IVR navigation, and DTMF capture built in.
- QA on every call, not a sample. Retell Assure, announced 17 December 2025 and available from 1 January 2026, scores 100% of calls on latency, interruptions, hallucinations, and sentiment. Human QA teams typically review 1-2% of calls; this is the feature that moves Retell from “demo tool” to “contact-center tooling.”
- Agents call your MCP servers. The MCP Node (July 2025) lets a voice agent call any MCP or HTTP service mid-call, so an internal order-lookup service you already expose over MCP is reachable without a bespoke connector. The direction matters: this is the agent using your tools, not an official Retell MCP server for querying call data from Claude.
- Compliance is self-serve. SOC 2 Type 1 and Type 2, HIPAA, and GDPR are documented, and the BAA and DPA can be signed online rather than negotiated, which makes it one of the few platforms a small clinic group can deploy for PHI without an enterprise contract.
Pricing reality
The list rate is $0.07–$0.31 per minute, and that range hides the decision. The pricing page breaks it down as $0.055/min voice infrastructure, $0.015/min for Retell’s platform voices (ElevenLabs voices run $0.04–$0.10), $0.015/min for Retell-supplied telephony (custom SIP is free), plus the LLM: $0.008/min for GPT 5 mini, $0.064/min for Claude 5 Sonnet or GPT 5.6 Terra, $0.16/min for GPT 5.5. A mainstream build — platform voice, Retell numbers, a mid-tier model — lands near $0.15 per minute all-in. A budget build on a small model lands near $0.09; premium voice plus a frontier model passes $0.30.
Then the add-ons: knowledge base +$0.005/min, PII removal +$0.01/min, and AI QA at $0.10/min after the first 100 free minutes. Running QA on every minute on that mainstream build raises the rate by roughly two-thirds — which is why most teams score a sample, not everything. Concurrency is 20 calls free, then $8 per concurrent call per month.
At 100,000 minutes a month, the mainstream build is about $15K/month, or roughly $180K a year, before QA. Enterprise pricing is custom, starts at 50+ concurrent calls, and is where volume discounts and dedicated implementation support live.
Best for
Support, CX-ops, and RevOps teams with at least one engineer who will own the agent like any other service — building inbound support lines, appointment scheduling, lead qualification, or outbound reminders at roughly 5,000 to 500,000 minutes a month, where per-minute pricing beats a six-figure enterprise commitment.
Skip it if nobody on your team can write a webhook or debug a tool call, if you need a vendor to design and run the agent for you, or if your contact center already runs NICE, Genesys, or Five9 and their native AI clears your containment bar.
Versus the alternatives
Vapi is the other default in developer voice platforms, and the better-funded one: a $50M Series B in May 2026 and over 1 billion calls handled. Its $0.05/min covers orchestration only — speech, LLM, voice, and telephony are billed separately — so it is cheaper to start and harder to forecast, and HIPAA is a paid add-on rather than self-signed. Pick Vapi when your engineers want to swap every component of the stack; pick Retell when you want one predictable meter and built-in QA.
ElevenLabs Agents is the fastest-growing entrant, riding ElevenLabs’ $500M+ ARR; pick it when voice quality and language coverage decide the project. Synthflow is the no-code option, now with enterprise contracts from about $30K a year — pick it when the builder must be a non-engineer. If you want a vendor to own delivery, buy managed instead: PolyAI for healthcare and hospitality systems of record, Parloa for EU financial services, Sierra when one agent must transact across chat and voice.
If none fit, run Retell’s $10 free credit against your single highest-volume intent before buying anything; a one-intent pilot tells you more than any vendor bake-off.
Watch-outs
- The headline rate is not your rate. $0.07 assumes the cheapest voice and excludes the LLM and telephony. Guard: price your exact build — model, voice, number source, add-ons — on the pricing page, then multiply by measured average handle time from a pilot, not by an estimate.
- Pay-as-you-go support is community and email. Reviewers repeatedly describe slow responses and account holds during identity verification, and a voice line that is blocked for a week is an outage. Guard: complete verification before launch, keep a human fallback number in your IVR, and move to an enterprise agreement with named support before the line carries revenue-critical calls.
- Latency figures are self-reported. Retell’s own March 2026 benchmark puts responses near 600ms; no independent benchmark confirms it, and a frontier LLM or premium voice adds to it. Guard: measure end-to-end latency on your chosen model and voice during the pilot, and set a threshold that triggers a model downgrade.
- The company is lightly capitalized for its scale. Roughly $5.1M raised against an estimated $80M revenue run rate means profitability is carrying the business, and a funding or acquisition event could change terms. Guard: keep agent prompts, flows, and tool schemas exported to your own repository via the API, and run telephony on your own Twilio or Telnyx SIP trunk so your numbers move with you.