Skip to content
Level up your prompts with SurePrompts — curated prompts for every workflow.
agentscamp

Best Voice Agent Platforms in 2026: Vapi vs Retell AI vs Bland AI vs LiveKit

Vapi, Retell AI, Bland AI, LiveKit Agents, Pipecat, ElevenLabs Agents and OpenAI Realtime compared on build-vs-buy, telephony, custom LLMs and open source.

By Imtiaz RayhanUpdated 11 min read

On this page
  1. The summary table
  2. Buy, build, or just the model
  3. Vapi: the developer-first buy
  4. Retell AI: the operations team's buy
  5. Bland AI: the enterprise stack on its own models
  6. LiveKit Agents: the open-source build
  7. Pipecat: the open-source pipeline
  8. ElevenLabs Agents (ElevenAgents): voices first
  9. OpenAI Realtime API: the model, not the platform
  10. How to choose
  11. Continue exploring

Vapi is the buy-side default when developers own the agent, Retell AI when an operations team does, and Bland AI when a regulated enterprise wants one vendor running its own models. Build on LiveKit Agents or Pipecat, both open source, when voice is the product. ElevenLabs Agents is voice-first; the OpenAI Realtime API is a model API, not a platform.

Key takeaways

  • Buy-side defaults: Vapi when developers own the agent, Retell AI when an operations team does, Bland AI when a regulated enterprise wants one vendor running its own models.
  • Build-side: LiveKit Agents (Apache-2.0) and Pipecat (BSD-2-Clause) are the open-source frameworks; you own the pipeline, the providers and the hosting, with LiveKit Cloud and Pipecat Cloud as optional managed paths.
  • Model freedom is the real divider: Vapi, Retell, LiveKit, Pipecat and ElevenLabs let you pick the LLM, and Retell and Vapi list Claude directly; Bland runs its own models and the OpenAI Realtime API serves OpenAI models only.
  • Every buy-side platform now has an MCP story: Vapi, Retell and Bland each ship an MCP server for coding agents, Retell and Vapi agents can call MCP tools mid-call, and ElevenLabs agents can be created from any MCP client.
  • All of these are usage-priced per minute (the OpenAI API per token); this roundup quotes no prices, and each tool page carries the dated figures.
  • Watch the dated deadlines: Bland requires production pathways and personas to move to its new agent builder by November 15, 2026, and OpenAI tells integrations built on the earlier Realtime interface to migrate to the GA one.

For most teams shipping a phone or web voice agent in 2026, the buy-side default is Vapi if developers own the project and Retell AI if an operations team does; Bland AI is the pick for regulated enterprises that want one vendor running its own models. When voice is the product and you want to own the pipeline, build on LiveKit Agents or Pipecat, the two open-source frameworks. ElevenLabs Agents (branded ElevenAgents on the site as of October 2026) is the platform to choose when its voices are the point; the OpenAI Realtime API is a model API that several of these platforms expose, not a platform.

Last reviewed: October 2026.

Every platform here is usage-priced per minute (the OpenAI API per token) and the numbers change often, so this roundup quotes no prices; see each tool page for dated figures. We quote no latency or market-share claims either; every vendor publishes its own.

The summary table

PlatformModelsBuild-vs-buy positionTelephonyCustom LLM / ClaudeOpen source?
VapiYour choice of STT, LLM and TTS providersBuy, developer-firstInbound and outbound numbers, bring your own carrierAny OpenAI-compatible endpoint; Anthropic among the providersNo (open SDKs)
Retell AIGPT, Claude, Gemini or speech-to-speech, per agentBuy, operations-firstRetell numbers, Twilio, Vonage, any SIP trunkClaude models listed; custom LLM over WebSocketNo (open SDKs)
Bland AIBland's own STT, LLM and TTSBuy, regulated enterpriseBland numbers, your Twilio, SIP trunksOwn LLM by default; custom models enterprise-onlyNo (open CLI and web SDK)
LiveKit AgentsPlugins for major providers, or LiveKit InferenceBuild, framework plus optional cloudFull telephony supportAny provider through pluginsYes, Apache-2.0
PipecatAny provider through pluginsBuild, Python pipeline frameworkNot stated in our sourcesAny LLM; Claude Code skills for scaffoldingYes, BSD-2-Clause
ElevenLabs Agents (ElevenAgents)In-house voices; LLM-agnosticBuy, voice-firstTelephony with Twilio and Genesys connectors, SIP trunkBring your own model, or hosted modelsNo
OpenAI Realtime APIOpenAI speech-to-speech modelsModel API, not a platformSIP through a trunking providerOpenAI models onlyNo

Buy, build, or just the model

The seven split into three postures, which the realtime voice agents guide calls build, assemble and buy. Vapi, Retell, Bland and ElevenLabs are platforms: they own turn-taking, interruptions, telephony and the operations tooling, and you configure agents. LiveKit Agents and Pipecat are frameworks: you write the agent in Python or Node, choose every provider, and run it yourself or on their clouds. The OpenAI Realtime API sits underneath both camps as a speech-to-speech model you can call directly.

What divides them is less the feature list than who picks the model and who is on the hook when a call goes wrong.

Vapi: the developer-first buy

Vapi describes itself as "the developer platform for building voice AI agents", and the design follows from that sentence. Every assistant combines three components, speech-to-text, an LLM and text-to-speech, and the docs promise "full control over each component, with dozens of providers and models to choose from". The LLM can be any OpenAI-compatible endpoint, which the docs list as "OpenRouter, AnyScale, Together AI, or your own server", and Anthropic is among the first-party providers. Assistants, Squads (multi-agent handoffs that keep context), tools and tests can be kept "as files in git", a config-as-code posture no other buy-side platform leads with.

MCP runs both ways: assistants connect to MCP servers for tools at runtime, and a Vapi MCP Server exposes the Vapi API to coding agents. The 2026 additions were Vapi Monitoring (April), simulations (August) and a Series B announced on May 12, 2026, led by Peak XV with YC, M12 and Kleiner Perkins.

Verdict: the buy-side pick for engineering teams. You get a live phone agent fast, keep model choice, and the agent lives in the repo rather than a dashboard. The trade is that operations features such as QA and campaigns are thinner than Retell's.

Retell AI: the operations team's buy

Retell is the same shape as Vapi with the center of gravity moved to the people who run phone operations. The pitch is a drag-and-drop canvas where you "add functions, integrations, call transfers, DTMF, SMS, and custom logic", then "simulate thousands of scenarios, and turn real production calls into regression tests". Around it sit contact-center features: IVR detection, live listen-in and takeover, outbound campaigns, a built-in CRM (June 2026), an AI QA Analyst (January 2026) and Conductor (July 2026), which simulates proposed changes before they ship.

Model choice is explicit. Models are grouped into "versatile and highly intelligent", "fast and cost-efficient" and "speech to speech", with the per-minute price next to each, and the dropdown lists the GPT 5.x line, Claude 5 Sonnet, Claude 4.6 Sonnet, Claude 4.5 Sonnet and Gemini 3.5 Flash. Your own LLM plugs in over a WebSocket, one connection per call: Retell handles telephony, transcription and turn-taking while your server generates each reply. The catch is documented plainly: "The LLM playground can't test a custom LLM agent, and simulation and batch testing reject them."

MCP runs both ways here too. Agents can call a remote MCP server's tools during a live call, flows have an MCP node, and Retell's own MCP server plus a Claude connector let you build, test and analyze agents from Cursor, Claude Desktop, Claude Code or a Claude conversation. Compliance is HIPAA, GDPR and SOC 2 Type 1 and 2.

Verdict: the buy-side pick when an operations team will own the agent after launch. The flow builder, testing and QA are the differentiators; keep the model inside Retell's list if you want the testing tools.

Bland AI: the enterprise stack on its own models

Bland makes the opposite bet to Vapi and Retell. Rather than orchestrating other vendors' models, it says it is "the only voice AI that runs entirely on its own models", with STT, LLM and TTS "built and run by one team", 40+ languages and real-time translation in 23 of them. Its pitch is regulated work: "Bland automates the calls where a single misstep can mean a lawsuit or a patient in danger", and it claims "230+ regulated companies". Enterprise customers get isolated, dedicated infrastructure, versioned Agent Releases, on-prem or VPC options and forward-deployed engineers. Compliance, per the pricing FAQ, covers SOC 2 Type I and II, HIPAA with a BAA, GDPR with a DPA, PCI DSS self-attested, FedRAMP 20x Class A and ISO 27001.

Two 2026 changes matter. Norm (March) is an assistant that "builds, edits, tests, and debugs your agents" from a plain-English description. The new dashboard (October 13) replaces pathways and personas with a single agent that has dev, staging and production environments, versions and branches, and Bland says all production pathways and personas must be migrated by November 15. The Bland MCP server (Streamable HTTP, OAuth sign-in from ChatGPT, Claude and Claude Code, or an API key) lets a coding agent place calls, build pathways, run evals and query analytics. Custom or fine-tuned LLMs are an enterprise conversation, not a self-serve setting.

Verdict: the pick when a regulated buyer wants one vendor accountable for every layer, and the wrong pick when you want to choose or swap the LLM yourself.

LiveKit Agents: the open-source build

LiveKit Agents is the build posture's default. "The Agents framework lets you add any Python or Node.js program to LiveKit rooms as full realtime participants", and the whole ecosystem is "open source under the Apache 2.0 license"; the agents repository had about 14,900 stars as of October 2026, with livekit-agents@1.8.8 released on 2026-10-15. You bring models through plugins "for most major AI providers", or use LiveKit Inference on LiveKit Cloud without separate API keys. The framework ships a custom turn-detection model, multi-agent handoff, and "built-in agent server orchestration, load balancing, and Kubernetes compatibility", plus "full support for telephony, so the user can join the call from a phone". Deploy to LiveKit Cloud or any custom environment; a no-code Agent Builder covers browser prototyping.

Verdict: the choice when voice is core product and you want control over every stage and the unit economics that come with running it. Budget real engineering time; the payoff is a stack you own.

Pipecat: the open-source pipeline

Pipecat is "an open-source Python framework for building real-time voice and multimodal conversational agents", BSD-2-Clause, "maintained by Daily and the community", with about 16,700 stars as of October 2026 and v1.13.0 released on 2026-10-19. It composes a pipeline from interchangeable STT, LLM and TTS pieces over WebSocket or WebRTC transports, supports "a full multi-agent system where specialists hand off, fan out in parallel, and coordinate over a shared bus", and ships client SDKs from JavaScript and React to Swift, Kotlin, C++ and ESP32. Around it sit pipecat init quickstart, Pipecat Flows, the Whisker debugger and Pipecat Cloud. It is also the most coding-agent-friendly entry: Pipecat Skills install into Claude Code with claude plugin marketplace add pipecat-ai/skills, and pipecat init lets Claude Code or Codex build the project for you.

Verdict: the build-side choice when the pipeline is your differentiation and you do not want to adopt a media-server worldview. It pairs with LiveKit or other transports underneath; the voice agent pipeline guide walks through the stages it orchestrates.

ElevenLabs Agents (ElevenAgents): voices first

ElevenLabs renamed Conversational AI to "ElevenLabs Agents" on 2025-09-03, and as of October 2026 its product page and docs brand it "ElevenAgents"; we found no dated announcement for that second name. The product promises agents "in 90+ languages" across "telephony, web, mobile, WhatsApp, SMS, email, and chat", with native connectors for Twilio, Genesys, Zendesk and Salesforce and a SIP trunk option. It is "LLM agnostic - bring your own model, call a frontier model via API, or use our hosted, domain-tunable models", reaches external systems "through MCP, APIs, and webhooks", and agents can be created and updated "from any MCP client". The voices are the draw: a library of over 16,000, cloning and design, and models built in-house. Compliance covers SOC 2 Type II, ISO 27001, PCI DSS Level 1, HIPAA and GDPR, with zero-retention mode and US, EU and India data residency.

Verdict: the buy-side pick when voice quality and multilingual reach are the product requirement, and a strong option when you already pay ElevenLabs for TTS. For voice-quality comparisons, see the TTS API roundup.

OpenAI Realtime API: the model, not the platform

The Realtime API is how you "build a speech-to-speech voice agent" on OpenAI's models directly; the documentation sample uses gpt-realtime-2.1, which takes text, audio and image in and returns text and audio. Sessions run "over WebRTC in the browser or WebSocket on the server", and a phone number can reach the API over SIP through "a SIP trunking provider (e.g., Twilio)". Tools come from "functions, MCP servers, and connectors". The API is generally available, and OpenAI tells integrations built on the earlier interface to migrate.

What it lacks is everything the platforms above sell: campaigns, QA, live monitoring, flow builders, regression tests. That is why Retell lists "GPT Realtime 2.1" and "GPT Realtime mini" as speech-to-speech options and Open WebUI's v0.12.0 realtime calls run on it.

Verdict: build on it directly if you want the fewest moving parts, OpenAI models and your own operations tooling; otherwise reach it through a platform.

How to choose

  • Developers own the agent and you want to ship this week: Vapi. Keep the assistant in git and swap models.
  • An operations team will run the agent after launch: Retell AI. Use the flow builder, simulation and QA, and stay on its model list.
  • A regulated buyer wants one accountable vendor and dedicated infrastructure: Bland AI. Accept that the LLM is Bland's, and plan the November 15 migration if you are on pathways today.
  • Voice is the product and you want to own the stack: LiveKit Agents for the full open-source platform with telephony, Pipecat for a pipeline framework that leaves transport choices open.
  • Voice quality and languages are the requirement: ElevenLabs Agents.
  • You want the fewest moving parts on OpenAI models: the Realtime API directly, with SIP for phones.
  • You need Claude specifically: Retell lists Claude models, Vapi mixes Anthropic with any STT and TTS, ElevenLabs is LLM-agnostic, and both open-source frameworks take it through plugins.

Whatever you choose, cost it with platform fee, model pass-through and telephony as separate line items; a headline per-minute number hides the comparison.

Continue exploring

Sources and further reading

Primary documentation used to verify this guide.

Frequently asked questions

What is the best voice agent platform in 2026?
For most teams buying rather than building, Vapi if developers own the project and Retell AI if an operations team does; Bland AI for regulated enterprises that want one vendor running its own speech and language models. If voice is your core product and you want to own the pipeline, build on LiveKit Agents or Pipecat, the two open-source frameworks.
Vapi vs Retell AI: which should I choose?
Both are hosted, model-agnostic orchestration layers with per-minute pricing, custom-LLM hooks and MCP in both directions. Choose Vapi for a developer workflow: any OpenAI-compatible endpoint as the LLM, assistants and tests kept as files in git, Squads for multi-agent handoffs. Choose Retell for an operations workflow: a drag-and-drop flow builder, simulation and batch testing, QA on production calls, campaigns and a built-in CRM.
Which voice agent platforms are open source?
LiveKit Agents is Apache-2.0 and Pipecat is BSD-2-Clause; both are frameworks you run yourself, with LiveKit Cloud and Pipecat Cloud as optional hosted paths. Vapi, Retell AI, Bland AI and ElevenLabs Agents are proprietary platforms with open SDKs, and the OpenAI Realtime API is a hosted model API.
Can I use Claude in a voice agent?
Yes on most of them. Retell lists Claude 5 Sonnet, Claude 4.6 Sonnet and Claude 4.5 Sonnet in its model dropdown; Vapi lets you mix Anthropic with any STT and TTS provider, or point it at any OpenAI-compatible endpoint; ElevenLabs Agents is LLM-agnostic; LiveKit Agents and Pipecat take any provider through plugins. Bland runs its own LLM, with custom models reserved for enterprise deals, and the OpenAI Realtime API serves OpenAI's speech-to-speech models only.
Do I need a platform, or can I build on the OpenAI Realtime API directly?
You can build directly: the Realtime API handles speech-to-speech over WebRTC in the browser or WebSocket on the server, takes phone calls over SIP through a trunking provider, and supports functions, MCP servers and connectors. What it does not give you is the operations layer, such as campaigns, live monitoring, QA, flow builders and regression tests, which is what Retell, Vapi and Bland sell, and several of them expose the same OpenAI Realtime models inside their platforms.

Filed under

comparison · voice · voice-agents · best-of

Related