Baseten Opens Voice Agent Stack and Realtime Voice in Private Preview

Unite.AI
Read full story· 4 min
Baseten launched the Voice Agent Stack and Baseten Realtime Voice in private preview on November 16, 2026, achieving about one-second median latency across 56 concurrent sessions. The stack co-locates speech models and orchestration for scalable, customizable voice AI pipelines with model choice and tuning options.

What happened

Baseten introduced the Voice Agent Stack and Baseten Realtime Voice in a private preview on November 16, 2026. The products are available to users through a waitlist and are designed to enable scalable, customizable voice AI pipelines. In testing, the system achieved about one-second median response time across 56 simultaneous sessions. The stack allows customers to select and tune individual speech models and orchestration components for their needs.

Key details

  • Private preview launched November 16, 2026, via waitlist.
  • System reached median time to first audio of 1.14 seconds at 56 concurrent sessions.
  • Estimated cost is $0.012 per audio minute at 32 sessions and $0.008 per call-minute at 56 sessions, excluding some model delegation costs.
  • Customers can independently select speech-to-text, language, and text-to-speech models, and switch components without redeployment.
  • API is compatible with OpenAI Realtime event protocol, with support for audio-in/audio-out and text-in/audio-out modes.

Why it matters

Companies and developers building voice AI applications can now use Baseten’s stack to run large numbers of voice interactions with low latency and customizable model choices. This enables more efficient scaling and cost control for operators, as they can adjust capacity at each pipeline stage and fine-tune models for specific use cases. The new stack also offers flexibility in integrating third-party models and adapting to industry-specific requirements, reducing the need for frequent redeployment or infrastructure changes.


More in Agents