Twilio

What it is

Twilio is the communications API company — phone numbers, voice, SMS, WhatsApp, email, and the carrier relationships behind them — that now sells itself as conversation infrastructure for AI agents. For a support team it shows up in three forms: the raw pipes under almost every voice-AI vendor (Vapi, Retell, and ElevenLabs all connect to Twilio numbers), ConversationRelay for teams that want to run their own voice agent directly on Twilio, and Flex, a programmable contact center you build rather than configure.

It is a large, profitable public company (NYSE: TWLO). Q2 2026 revenue was $1.50B, up 22% year over year, with voice growing above 20% and self-serve voice above 50% — growth Twilio ties to AI-native builders. In November 2025 it bought Stytch, an identity platform, for $104.1M, to authenticate humans and AI agents on the same rails.

What changed in 2026

• A conversation layer above the pipes. At SIGNAL on 6 May 2026 Twilio made four products generally available: Conversation Orchestrator (one thread across channels, with handoffs between AI and human agents), Conversation Memory (an identity-resolved customer profile built for LLM recall), Conversation Intelligence (LLM “language operators” that flag sentiment shifts and escalations mid-conversation), and Enterprise Knowledge (grounding on your FAQs and policies). Agent Connect, an open-source toolkit, plugs agents built on any LLM or framework into these.
• ConversationRelay grew up. Twilio’s voice-AI bridge handles speech-to-text, text-to-speech, and turn-taking while your server runs the LLM over a WebSocket. In 2026 it added HIPAA eligibility, PCI compliance, Deepgram Flux turn detection, and Conversation Relay Insights (GA 6 May 2026), a dashboard for latency, interruptions, and handle time per call.
• Flex modernized and pruned. The Flex SDK went GA on 26 March 2026, and a $35 per monthly active user plan arrived for Flex on existing Twilio accounts. Programmable Chat in Flex reached end of life on 1 June 2026 (migrate to Flex Conversations), and Frontline, the mobile app for field reps, retired on 30 September 2026.
• MCP. The @twilio-alpha/mcp server exposes Twilio’s public API to Claude or another client, filterable by service with --services and --tags. It is still branded Alpha.

Pricing reality

The API layer is pay-as-you-go with no platform fee. US voice is $0.0085/min inbound on a local number, $0.022/min inbound toll-free, $0.014/min outbound, and $0.004/min over SIP; numbers cost $1.15/month local and $2.15/month toll-free. ConversationRelay adds $0.07/min. A self-built voice agent therefore runs about $0.08/min before the LLM — 100,000 inbound minutes a month is roughly $7,850 in Twilio charges plus whatever your model provider bills. That undercuts a mainstream Vapi or Retell build ($0.09–$0.15/min all-in) only if your own engineers absorb the orchestration work those vendors sell.

The conversation layer meters separately: Memory at $0.005 per daily utilized profile plus $0.007 per recall, Intelligence at $0.005 per 1K characters for Twilio-authored operators, Orchestrator at $0.0002 per 1K characters ingested, Knowledge at $0.005 per retrieval. At 100,000 AI calls a month with one recall each and about 10,000 profiles read a day, Memory alone is roughly $2,200 a month.

Flex has three plans that cannot be mixed: $35 per monthly active user plus usage, $150 per named user per month on contract, or $1 per active user hour. A 50-agent voice team on named users is $90K a year before telephony — close to Five9’s $95K Core floor — but the build is yours. Third-party estimates put outside implementation help at $10,000 and up; treat that as a floor for anything beyond a basic queue. Agent Copilot (beta) adds $0.035 per voice minute.

Best for

Support-engineering and CX-ops teams with in-house developers who want to own their voice and messaging stack — either running an AI agent on ConversationRelay with their own LLM, or building a custom contact center on Flex where the agent desktop, routing, and CRM integration are code they control. It is also the default telephony layer underneath Vapi, Retell, or ElevenLabs when you want your numbers on your own account so they move if you switch vendors.

Skip Flex if nobody will maintain it after go-live: a packaged CCaaS ships workforce management, quality management, and reporting that Flex makes you build or buy from partners.

Versus the alternatives

Genesys Cloud and Five9 are the market-share leaders in packaged CCaaS. Pick them over Flex when you need WFM, QA, and supervisor reporting on day one and your team configures rather than codes. Pick Flex when your routing or agent desktop is unusual enough that a packaged suite would need heavy customization anyway.

Vapi and Retell AI sit on top of Twilio for voice agents. Pick them over raw ConversationRelay when you want testing, call analytics, and component swapping without writing the orchestration yourself; pick ConversationRelay when you already run a Twilio account at volume and want one vendor, one BAA, and one bill.

ElevenLabs Agents is the fastest-growing entrant in voice agents — pick it when voice quality and language coverage decide the project, and keep Twilio as the carrier underneath. Telnyx is the cheaper-carrier alternative for the pipes alone (about $0.0055/min inbound on Vapi’s pricing calculator).

If none fit, run one high-volume intent on a Twilio number through Vapi for a month, then rebuild the same intent on ConversationRelay and compare cost per resolved call. That tells you whether the orchestration vendor earns its margin.

Watch-outs

• Five or six meters, not one. Telephony, ConversationRelay, LLM, Memory, Intelligence, and carrier fees each move independently — Twilio said incremental US carrier fees added about 10 points to messaging growth in Q2 2026. Guard: tag every call and message with a cost-centre attribute in your own logs and reconcile against the Twilio usage API monthly, not against the pricing page.
• Flex is a build, not a purchase. Reviews on G2 cite complex setup and thin built-in reporting. Guard: scope WFM, QA, and reporting before signing — either budget engineering time for them or price the partner add-ons — and keep a named owner for the Flex codebase after go-live.
• Products get retired. Programmable Chat in Flex and Frontline both reached end of life in 2026, after standalone Programmable Chat was deprecated. Guard: subscribe to the Twilio changelog, build on the currently recommended surface (Flex Conversations, ConversationRelay, the Conversations APIs), and keep agent logic in your own service so a deprecation changes an adapter, not your agent.
• The MCP server is Alpha. Guard: give it a restricted API key scoped to the services you need, use --services to limit what it exposes, and do not run it alongside community Twilio MCP servers, as Twilio’s own security team advises.