Voice AI platform for building AI phone agents that handle calls end to end
Retell AI is a voice AI platform for building and deploying AI phone agents that handle inbound and outbound calls.
AI Panel Score
6 AI reviews
Reviewed
Facts checked against retellai.com
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.With Retell AI, teams build phone agents that answer or place calls, hold two-way conversations, and complete tasks like after-hours support, appointment reminders, lead qualification, collections, and order status checks. Agents run on a telephony layer that includes phone number provisioning, SIP trunking, call transfers with warm handoffs, and call recording, and can use a customer's own Twilio account instead of Retell-provisioned numbers. During a call, agents can read and write to CRM, helpdesk, and calendar systems and call internal functions or webhooks to act on live data.
The platform supports multiple LLMs, with the model choice made per agent based on call complexity and swappable without rebuilding the agent. Developers can also connect a custom LLM over WebSocket while Retell handles telephony and turn-taking, or connect a live agent to remote MCP servers so it can call their tools mid-call. A Retell MCP server (hosted at mcp.retellai.com, also available as the @retell-ai/mcp-server npm package) lets MCP-capable clients such as Cursor, Claude Desktop, and Claude Code build and manage Retell agents directly. Additional integrations include official nodes for n8n, a Zapier connection, and two-way CRM sync with Salesforce and HubSpot. A CLI is available for managing agents, phone numbers, and knowledge bases from the terminal.
Retell AI is built for customer service and inbound support teams, operations teams who build call flows on the drag-and-drop canvas, developers who want direct access to voice and orchestration primitives, and enterprises that need compliance controls and call volume at scale. Pricing is usage-based: $0.07-$0.31 per minute for voice agents and from $0.002 per message for chat agents, with no platform fee and $10 in free credits at signup. An enterprise plan is available for high-volume buyers requiring custom deployment, volume pricing, higher concurrency, and dedicated support through direct sales contact rather than self-serve signup.
The platform is API-first, with SDKs for Node.js and Python, an API reference for authentication and endpoints, and documentation covering deployment from prototype to production. It holds SOC 2 Type I and Type II certification and an ISO 27001:2022 certificate, and supports HIPAA compliance with a self-signed BAA, GDPR compliance, configurable per-agent data retention for call and chat data, and KYC verification required to unlock outbound calling, phone number purchases, and SMS.
Supports multiple LLMs and lets teams select or swap models per agent based on call complexity without rebuilding the agent.
Supports conversations across 50+ languages, including switching languages mid-call.
Exposes API primitives so engineering teams directly control routing, prompts, and integrations alongside a drag-and-drop canvas for designing call flows visually.
Records calls for review and compliance purposes.
Transfers calls to human agents with warm handoff support during a live conversation.
Allows the agent to stop mid-sentence when a caller interrupts, enabling natural back-and-forth dialogue.
Targets ~600ms response latency so AI phone agents can hold real-time conversations with callers.
Provisions phone numbers directly for deploying voice agents.
One-toggle prompt presets for personality and tone (Default Tone, Natural Filler Words, High Empathy), accuracy (Echo Verification, NATO phonetic alphabet, speech normalization) and trust and safety; new agents start with Default Tone and AI disclosure when asked.
Runs a customer's own LLM behind a Retell voice agent over a WebSocket while Retell manages telephony and turn-taking.
Adds emotion tags (empathetic, excited, happy, curious, surprised), effects such as sighs and pauses, and word emphasis to agent speech on Retell platform voices, from an emotion palette you choose or tags written into the prompt.
Clones a custom voice for the agent to maintain a consistent brand voice across calls.
Reads and writes to CRM, helpdesk, and calendar systems in real time during a call, with two-way contact sync and call logging for Salesforce and HubSpot.
Connects a live voice agent to remote MCP servers so it can call their tools mid-call, and lets developers build and manage agents from MCP-capable clients like Cursor and Claude.
Connects agents to telephony infrastructure via SIP trunking, including support for custom providers like Twilio, Telnyx, and Vonage.
Connects agents to internal systems and tools via webhooks and function calling during a call.
Provides official nodes for n8n and a Zapier connection to link Retell agents with thousands of external apps and workflows.
Self-serve usage pricing of $0.07-$0.31 per minute for voice agents and from $0.002 per message for chat agents, with no platform fee; sign up in the dashboard and get an API key with $10 in free credits, no sales contact required.
For high-volume and enterprise buyers requiring custom deployment, volume pricing, higher concurrency, security terms, and dedicated support; requires talking to sales instead of self-serve signup.
API-first phone agents with real telephony depth, priced per minute with no platform fee.
“Retell pairs a drag-and-drop flow builder with direct API control over voice agents. The telephony and compliance layer looks production-ready, not a demo toy.”
A quoted ~600ms latency, mid-sentence interruption handling, and 50+ languages with mid-call switching. The latency number times one stage, not a full turn; the independent Cekura tests Retell publishes put a complete turn near two seconds, yet still rank it first on repeatable reliability. That's a spec sheet with receipts, not marketing fog. Warm handoffs and SIP trunking with your own Twilio number mean you're not locked into their telephony from day one.
Per-minute pricing with no platform fee, self-serve with a free trial through the dashboard. Enterprise tier is quote-only for volume and concurrency, which is normal at that scale. SOC 2 Type I and II, HIPAA, GDPR — the compliance boxes for regulated call centers are checked.
The tradeoff: cost predictability. The rate runs $0.07 to $0.31 a minute depending on the LLM and voice you pick, so model choice is a budget call as much as a quality one. Non-engineers can lay out call flows on the drag-and-drop canvas, but the real depth sits in the API, SDKs and webhooks. For anyone shipping real phone automation, that's the right tradeoff.
Custom LLM over WebSocket and remote MCP tool-calling mid-call are architecture choices closed builders don't expose.
SOC 2, HIPAA, GDPR support plus bring-your-own-Twilio reduces lock-in, though KYC gating outbound calling adds friction.
Self-serve signup with a free trial and per-minute pricing means a working agent can ship fast, without a sales call.
API-first primitives and MCP server support fit teams building real automation, not just cutting call-center costs.
Docs, blog, changelog, and a CLI plus SDKs for Node and Python point to active engineering investment.
Engineering teams building production phone agents who want API control on top of a visual flow builder.
Skip it if finance needs one fixed monthly number; per-minute costs move with the model and voice you choose.
Retell's tone direction and publish-time review are real, but I couldn't find one brand-voice standard across agents.
“Agent Handbook tone presets, Expressive Mode palettes and a version diff at publish give a Creative Director real control over how one agent sounds. Holding one brand voice across many agents and markets is still per-agent work.”
Three Personality & Tone presets sit in the Agent Handbook, and new agents start on Default Tone (Professional). Expressive Mode goes further: you pick the emotion palette, edit the override prompt deciding when empathy or a pause lands, and hand-place tags. Custom pronunciation takes IPA, CMU, Pinyin or Jyutping by provider, so brand names survive Mandarin and Cantonese.
Review has a home too. Version comparison opens in the publish modal, diffing prompt, voice and flow changes before release, and A/B Testing splits live calls to trial a new voice. For one agent, that's the approval loop a Creative Director wants.
The catch is the fleet: presets, palettes and voices are set per agent, and I couldn't find a shared brand-voice standard they inherit. Expressive tags work in 13 of the 50+ supported languages, so other markets get words without direction. Over three years, one brand voice rides on reviewing every agent's diffs, not a style guide.
Positions itself as infrastructure for phone agents rather than a closed builder, a clear stance in a fast-moving lane.
Creative can set tone, emotion palettes and pronunciation, then check voice changes in a version diff before publishing, though every setting lives on a single agent.
n8n nodes, Zapier, Salesforce and HubSpot two-way sync, plus MCP server access from Cursor and Claude Desktop, is a genuinely wide surface.
No platform fee keeps switching costs low, but voice settings live per agent and Expressive tags cover 13 languages, so multi-market consistency takes ongoing review.
Agent Handbook tone presets, Expressive Mode palettes with an editable override prompt, and IPA, CMU, Pinyin or Jyutping pronunciation show real voice craft.
Creative teams who want sign-off on how each phone agent sounds.
Brands that need one voice standard inherited by every agent in every market.
20,000 minutes runs $1,400 to $6,200 a month, and every component is priced in public.
“Pay-as-you-go runs $0.07-$0.31 a minute, with no platform fee, no contract and per-second billing. Budget from the component rate card, not the "$0.07 flat" line on Retell's Vapi comparison.”
Retell asks for no minimums and no contract. Then it publishes the whole rate card. Retell Voice Infra is $0.055 a minute. Most voices add $0.015. US telephony adds $0.015. LLMs run $0.0016 to $0.32 on standard tier. Billing runs to the second, no per-call rounding.
Take 5,000 calls a month at 4 minutes: 20,000 minutes. At $0.07, that's $1,400/month. At $0.31, $6,200. Year 1 lands between $16.8K and $74.4K. Silence and hold time bill too. Past 20 concurrent calls, add $8 per call per month.
The catch is Retell's own Vapi comparison. It faults Vapi for a platform fee, then quotes "$0.07 per minute flat," branded calls included. The rate card charges $0.10 per Branded Call and $0.01 a minute for PII Removal. A standard BAA is self-signed online on pay-as-you-go; a custom BAA takes the Enterprise plan. Budget from the rate card.
$10 in signup credits and invoices in the Billing tab each cycle; KYC must clear before outbound calling.
Pay-as-you-go carries no minimums and no contract; Enterprise terms are custom MSA/DPA, and I couldn't find term lengths.
Every component is priced in public, down to $0.005 add-ons; the Vapi comparison's "$0.07 flat" line is the one blur.
Per-second billing, no charge for failed calls and an AI fee that stops at transfer tie cost to handled minutes.
20,000 minutes a month spans $1,400-$6,200 on model and voice picks; concurrency past 20 calls adds $8 each monthly.
Finance teams who model voice costs from a per-minute rate card.
Buyers who need one flat per-minute price before approving a pilot.
Retell documents where its denoising breaks, but Expressive Mode only directs platform voices.
“Retell's audio settings come with real specs: three denoise modes, clone-recording guidance, and PCMU, PCMA or G.722 on SIP trunks. The catch for a producer is that emotion tags only run on platform voices, and Retell's published Cekura tests put its voice tone at 4.36 against ElevenLabs' 4.47.”
Remove noise is the default, and the docs say what it leaves alone: loud background speech. Step up to Remove noise + background speech (+$0.005/min) and Retell keeps only the dominant near-mic voice. The docs list where that breaks: speakerphone, a quiet or distant caller, IVR prompts, voicemail greetings, hold music. That's a mix engineer's warning label. Recordings keep the raw pre-cancellation caller audio, so the source survives.
Clone guidance reads like a session brief. One speaker, no music or echo, conversational speech, a clean mic over phone capture. I couldn't find a minimum length or sample rate, though. SIP trunks carry PCMU, PCMA or G.722.
Expressive Mode is the directing tool: pick the palette, write [pause] or [emphasis] into a line. The catch is platform voices only, so an ElevenLabs or Cartesia clone can't take the tags. Retell's published Cekura tests score its voice tone 4.36, ElevenLabs 4.47.
The daily fight is picking a denoise mode: the default leaves loud background speech in, and the aggressive mode can drop a short 'yes' on already-clean audio.
The denoise page lists where the aggressive mode fails, from speakerphone and distant callers to IVR prompts and hold music, and clone guidance covers recording conditions.
Self-serve signup with $10 in free credits and per-minute billing lets you audition voices and denoise modes on real calls before any contract.
Expressive Mode lets you pick the emotion palette, override when tags fire and write [pause] or [emphasis] into a line, though only on platform voices.
SIP trunking with PCMU, PCMA or G.722 over SRTP drops into an existing phone stack, and raw pre-cancellation recordings give a sound reviewer the real line.
Voice teams who tune caller audio for noisy real-world phone lines.
Producers who need directed emotional delivery from an ElevenLabs or Cartesia voice clone.
A phone agent platform the ops lead can actually run after launch, though full QA costs extra.
“Retell puts a no-code studio and a drag-and-drop call-flow canvas in the hands of whoever runs the agents after launch. The day-two tooling is real, though AI QA past the first 100 free minutes costs $0.10 a minute.”
The homepage pitch is aimed at whoever babysits these agents after launch, not just whoever builds them. A no-code studio "your team can operate," call flows on a drag-and-drop canvas, updates without relying on Retell for every change. Retell's own Vapi comparison page leans on that gap. Good. The ops lead lives in this at month three.
After go-live it earns its keep. Real production calls turn into regression tests, QA covers up to 100% of calls, and custom analytics compare A/B results per agent version. Agentic Human Handoffs and live listen-in cover the call that goes sideways. Not demo glow. The catch: AI QA is free for the first 100 minutes, then $0.10 a minute on top of the call, so checking every call is a budget decision.
I couldn't find a mobile app. For 2am calls, checking one live from your phone stays a question mark.
QA across calls, per-version A/B analytics and live listen-in show care for the daily operator, though I couldn't see empty states or micro-copy up close.
The drag-and-drop canvas and Pre-built Agent Templates give non-engineers a way in, with APIs and SDKs there when you need to go beyond the dashboard.
I couldn't find a mobile app, which matters less for a phone platform run from a desk but leaves on-the-go monitoring of live calls unclear.
Self-serve signup with $10 in free credits and Pre-built Agent Templates gets you moving fast, though KYC verification gates outbound calling.
Simulation testing, regression tests built from real production calls, and live listen-in with instant takeover point to a team that's thought about failure modes.
Support ops teams who want to tune phone agents without an engineer on every change.
Avoid if you want to QA every call on a fixed monthly budget.
Retell's 95.7% benchmark holds up, but its own latency data undercuts the homepage's 'leader in responsiveness' line.
“Retell names its benchmark source, its rivals and its per-minute rate, and it prints where it loses. The homepage hasn't caught up: its latency, '#1' and 'ahead of 6' lines each claim more than Retell's own pages back.”
Retell's blog is more careful than its homepage. The '95.7% workflow accuracy' line traces to Cekura Experiment 02: one Medicare agent on six platforms, 414 calls, Retell passing 22 of 23 evaluators, ElevenLabs next at 21. Named rivals, linked run reports. Fair. Nit: six platforms including Retell is five rivals, not 'ahead of 6.'
The latency line is the real tell. Homepage: 'Independent benchmarks confirm Retell as the leader in responsiveness,' at ~600ms. Retell's own post: 2.21s mean response, behind ElevenLabs at 1.27s, and 'Retell is not the fastest.' Same company, two stories. '#1' in the H1 cites nothing either.
What holds up: $0.07-$0.31 a minute on the pricing page, bring-your-own Twilio, and the Retell MCP server for Claude Code. But compliance wording drifts. 'Fully compliant' with ISO 27001 on the homepage, 'HIPAA-ready' on pricing. The trust center lists the SOC 2 reports and an ISO 27001:2022 certificate. Get them before you sign.
Leads Cekura's repeat-run reliability at 75.61% pass³ against Vapi's 59.76%, though ElevenLabs answers faster and sounds better.
SIP trunking and bring-your-own-Twilio soften lock-in, but agent logic and prompts still live in Retell's orchestration layer.
Docs, API reference, changelog and a status page are all public, and the benchmark post was updated in September 2026.
The benchmark post names its source and concedes losses, but the homepage's latency lead, '#1' and 'ahead of 6' lines outrun it.
The 95.7% traces to 22 of 23 evaluators in Cekura Experiment 02, and the same testing scores interruption handling 5.00 out of 5.
Engineering teams who want a vendor that publishes its benchmark losses.
You need the fastest per-turn response time in the category.
Common questions answered by our AI research team
Yes. Retell supports integration with any telephony system through SIP trunking, letting you connect your existing phone number and deploy an AI-powered answering service or AI IVR without changing providers.
Retell quotes ~600ms response latency. Independent Cekura testing that Retell publishes measures a full conversational turn at about 2 seconds on platform defaults, because the two figures time different parts of the call; in that same testing Retell scored 5.00 out of 5 on interruption handling.
Yes. Retell's compliance docs describe it as HIPAA and GDPR compliant and SOC 2 Type 1 and Type 2 certified, and its trust center lists an ISO 27001:2022 certificate. A signed BAA is required before you send protected health information through Retell, and it can be self-signed online.
Yes. You can manually test changes, simulate thousands of scenarios, and turn real production calls into regression tests to catch failures and edge cases before deployment.
Yes. Agentic Human Handoffs let AI agents brief a human agent and confirm readiness before connecting the customer, and you can also listen live and take over the conversation instantly whenever needed.
Company
Retell AI Inc.Founded
2023Pricing
Usage-basedFree Trial
Available
Retell AI, based in San Francisco, provides a platform for building and deploying voice AI agents for automated inbound and outbound phone calls.
