News
AssemblyAI’s comparison says the four voice-agent stacks sit at different layers, with trade-offs in control, transport and STT integration.
AssemblyAI’s comparison argues that Vapi, LiveKit, Retell and Pipecat are not direct substitutes, but different layers in a production voice-agent stack, with the main choice being how much of the pipeline a team wants to own.
Evidence and context
In the source article published Oct. 13, 2026, AssemblyAI describes Retell as a managed, low-code phone-agent platform; Vapi as an API-first orchestrator; LiveKit as open-source WebRTC transport plus an agent framework; and Pipecat as an open-source Python pipeline that developers assemble themselves. The article says all four depend on the underlying speech-to-text model, and that two of the products are often used together rather than treated as rivals on a single feature axis.
The comparison separates the products by orchestration, transport and transcription control. Retell and Vapi manage the conversation loop for the user, with Retell emphasizing visual builder workflows and built-in telephony, and Vapi exposing the pipeline through code and API configuration. LiveKit uses an event-driven WebRTC model in which the agent joins a room as a participant, while Pipecat leaves the pipeline explicit in code so developers can insert logic between steps. The article says Pipecat is the only one of the four where every step is directly inspectable and modifiable.
On transport, the article says Retell handles telephony and call transfer, Vapi abstracts telephony and SIP setup, LiveKit runs on WebRTC with an SFU, and Pipecat is transport-agnostic. It also notes that Pipecat can run over LiveKit transport, underscoring that the products are not mutually exclusive in deployment.
The piece gives public pricing examples as of Oct. 2026: Retell is listed as starting from $0.07 per minute; Vapi lists $0.05 per minute for hosting plus provider costs; LiveKit Cloud has a free tier with 1,000 agent minutes; and Pipecat is a free framework, with infrastructure and model costs borne separately. Those figures are presented as published pricing, not as comparative total cost of ownership.
For speech recognition, the article says Vapi exposes transcriber selection in assistant configuration, LiveKit uses plugins, Retell offers selectable transcription options and Pipecat provides direct API access. AssemblyAI says its Universal-3.6 Pro Realtime model, released Sept. 29, 2026, is available through integrations with Vapi, LiveKit, Pipecat and Retell. The article cites AssemblyAI’s voice-agent benchmark at 5.19% word error rate for Universal-3.6 Pro Realtime, versus 7.78% for ElevenLabs Scribe v2 and 8.64% for Deepgram Nova-3, and says the model posts a 14.4% entity error rate versus 26.1% for Nova-3. On Pipecat’s open STT benchmark, the article says it reached a 0.96% pooled semantic word error rate and a 307 ms median time to first speech.
The article also points to Vapi-specific product news from the supplied enrichment. Vapi announced GPT-Live beta for all customers on Oct. 2, 2026, adding a speech-to-speech architecture alongside its modular pipeline, according to Vapi’s own product announcement. The enrichment also says Vapi has documented support for SIP, Twilio, Telnyx and Plivo, and that its modular architecture lets developers select providers independently or supply custom model infrastructure.
AssemblyAI’s broader conclusion is that the “best” choice depends less on surface features than on whether a team wants managed speed or full control. Retell and Vapi reduce operational work. LiveKit and Pipecat increase flexibility, but shift responsibility for infrastructure, monitoring and scaling onto the developer. The article also highlights a fifth path: AssemblyAI’s Voice Agent API, which it says combines STT, LLM, voice generation, turn detection and interruption handling in a single WebSocket-based service at $4.50 per hour, with no separate orchestration layer.
The source includes one material limitation: many of the performance and pricing statements are vendor-specific and should be read as the publishers’ own figures or product claims rather than independent verification. The article also notes that interruption behavior depends on the orchestrator as well as the transcription model, which means speech quality alone does not determine production performance.
Be the first to comment