A Guth Labs publication ↗
Back to News

Baseten previews a voice-agent stack that brings models and orchestration together

Baseten previews a voice-agent stack that brings models and orchestration together

By guthlabs.newsroom

Baseten announced its Voice Agent Stack and Realtime Voice on November 16. Both are in private preview, not general availability. The stack hosts an agent's orchestration alongside its speech and language models, while letting components scale separately. Baseten announcement

The reference agent combines open-weight speech models with a Pipecat-based orchestrator and can delegate reasoning and tool work to another model. Baseten says its API follows the core OpenAI Realtime protocol. That is a compatibility claim, not a guarantee that every existing application will migrate unchanged.

The company's latency chart reports median time to first audio rising from 0.94 to 1.14 seconds as concurrent calls increase from one to 56. Those results depend on its configuration and context length, not an unconditional production guarantee.

Cost and quality claims require particular care. Baseten's reported 73.5% result on τ²-Bench Voice is self-reported, and its footnotes explicitly say it is not directly comparable with Artificial Analysis results. The associated dedicated-infrastructure cost calculation assumes full utilization and excludes per-token reasoning delegation. Results and footnotes

Guth News analysis: Co-locating the orchestrator addresses an easily missed part of voice latency: the network travel between components. Whether dedicated capacity is economical depends on traffic and utilization. Preview access and workload-specific testing matter more than treating a low per-minute figure as a universal price.

Sources and citations

Revision history

Published .