Deepgram Moves Into Voice Agents

Deepgram’s Voice Agent API shows that low latency speech infrastructure was no longer enough on its own, because the biggest revenue pool shifted to the orchestration layer that turns speech models into working phone agents. Instead of only selling transcription into Vapi and Retell, Deepgram started selling the full call loop, speech recognition, model routing, text to speech, barge in, and turn taking, as one API priced at $4.50 per hour, or $0.075 per minute.1 2 3 4

• Vapi and Retell were built as the neutral control layer above providers like Deepgram, OpenAI, and ElevenLabs. They let developers swap components, connect business systems, test flows, and monitor calls, while charging a platform fee on top of pass through model costs. Deepgram moving into agents attacks that control point directly.5 6 7
• The product move was concrete, not just branding. Deepgram added a unified voice to voice API with bring your own LLM and TTS support, then kept expanding it through AWS Bedrock integration and multi provider fallback. That turned Deepgram from a single speech model vendor into an opinionated runtime for production voice calls.2 3 8 9
• The OfOne acquisition in January 2026 shows where this goes next. Once a company owns the full agent stack, it can package it into vertical applications. Deepgram used a quick service restaurant ordering startup as the first wedge, because restaurant calls are repetitive, latency sensitive, and easy to price per minute or per order.4 10

The next step is a split market. Neutral platforms like Vapi and Retell keep winning teams that want model choice and workflow control, while Deepgram pushes toward bundled, opinionated agents and then into full vertical products. In voice AI, infrastructure vendors are increasingly trying to own the application surface where more of the margin and customer lock in sit.1 4 5 6 7