AssemblyAI's Voice Agent API: $4.50 an hour for the whole pipeline
AssemblyAI shipped its Voice Agent API on April 29 and the pricing is the headline. $4.50 per hour for the complete voice agent pipeline — speech-to-text, LLM reasoning, voice generation, all behind a single WebSocket. They built every layer on their own models top to bottom.
This is a direct shot at the Vapi / Pipecat / LiveKit + multi-vendor stitching stack that's been the default for the past year. AssemblyAI's argument: 76% of voice agent builders said STT accuracy is the single non-negotiable, and stitching three vendors together never gets the foundation right. Their answer is to fold STT (Universal-3 Pro Streaming, 99+ languages, real-time diarization, code-switching), the LLM, and TTS into one endpoint where the listening half is actually accurate.
Engineering details that matter for production: server-side turn detection that distinguishes natural pauses from completed turns, immediate interruption handling, configurable conversational feel, tool calling registered with JSON Schema, live config updates mid-conversation without reconnect, session resumption within 30 seconds after disconnect. Standard JSON over WebSocket, no proprietary SDK — drops straight into Claude Code or anything else.
The use cases they're targeting: contact centers, clinical intake, sales coaching, language learning. Translated: replacing humans on the phone, where every dollar of margin matters and $4.50/hour all-in is the number that lands the deal versus $15-20 stitched together.
Page: https://www.assemblyai.com/blog/introducing-our-voice-agent-api
← Back to all articles
This is a direct shot at the Vapi / Pipecat / LiveKit + multi-vendor stitching stack that's been the default for the past year. AssemblyAI's argument: 76% of voice agent builders said STT accuracy is the single non-negotiable, and stitching three vendors together never gets the foundation right. Their answer is to fold STT (Universal-3 Pro Streaming, 99+ languages, real-time diarization, code-switching), the LLM, and TTS into one endpoint where the listening half is actually accurate.
Engineering details that matter for production: server-side turn detection that distinguishes natural pauses from completed turns, immediate interruption handling, configurable conversational feel, tool calling registered with JSON Schema, live config updates mid-conversation without reconnect, session resumption within 30 seconds after disconnect. Standard JSON over WebSocket, no proprietary SDK — drops straight into Claude Code or anything else.
The use cases they're targeting: contact centers, clinical intake, sales coaching, language learning. Translated: replacing humans on the phone, where every dollar of margin matters and $4.50/hour all-in is the number that lands the deal versus $15-20 stitched together.
Page: https://www.assemblyai.com/blog/introducing-our-voice-agent-api
Comments