AssemblyAI shifts to voice agent runtime

This shifts AssemblyAI from being one component in a voice stack to owning the live conversation loop. Instead of developers wiring three vendors together, then debugging timing, interruptions, and tool calls across multiple streams, one WebSocket handles speech in, reasoning, and speech out. That makes the product feel less like a transcription API and more like a real time runtime for phone agents, assistants, and support bots.1 2 3 4

• The hard part is not just calling STT, an LLM, and TTS. It is making turn taking feel natural. AssemblyAI bakes in interruption handling, turn detection, session resume, tool calling, and noise suppression, which are the pieces developers usually end up coordinating themselves when they stitch vendors together.1 2 4
• This also changes who AssemblyAI competes with. The closest comparison is no longer only speech APIs like Deepgram, but voice agent platforms like Retell AI, Vapi, and Bland AI that package multiple model providers behind one developer surface. AssemblyAI is moving from supplier to full stack alternative.5 6 7
• The business upside is higher spend per customer. A pure transcription vendor gets paid for audio converted to text. A voice agent runtime can charge for the whole active conversation, and AssemblyAI now prices Voice Agent API at $4.50 per agent hour while company revenue reached an estimated $76M run rate by October 2026.2 3 8

The next step is deeper control of the entire call path, including telephony, custom tools, and model quality tuned for messy business conversations. As speech models commoditize, the winner will be the platform that makes a live agent feel reliable in production, not the one with the best standalone STT benchmark.2 4 6 7