Yelp and Hatch switch their voice AI to OpenAI's GPT-Live-1, a model built to handle interruptions and talk over background noise

OpenAI launched GPT-Live-1, a full-duplex voice model, in its API on September 10 -- letting a voice agent listen and speak at the same time instead of waiting for a caller to finish a sentence. The same day, Yelp rolled it into Yelp Host, its restaurant reservation line, and Hatch, its lead-management platform for service businesses -- the first announced production deployments outside OpenAI's own ChatGPT Voice.

The model decides several times a second whether to keep listening, pause, interrupt, speak, or hand off to a backend tool. OpenAI's own published benchmarks show it closing much of the gap that made earlier voice AI feel like a walkie-talkie conversation, one side waiting for the other to finish before responding: full-duplex interactivity climbs from 45.4% to 80.1%, turn-taking latency drops from 1.4 seconds to 0.8, and tool-calling accuracy rises from 60% to 87%.

Yelp Host has handled more than a million calls since it launched in October 2025, mostly on turn-based voice models that had to wait for a pause before responding. With GPT-Live-1, a caller can interrupt to add a dietary restriction, change party size mid-sentence, or talk to someone else in the room, and the system is meant to track the thread rather than restart it. Yelp's product chief said every call is now "more conversational and responsive"; the company has not published a number for how much call-transfer rates actually dropped, only that internal production testing showed an improvement.

Hatch, which manages leads for service businesses like plumbers and electricians across voice, text, email, and web, layers its own scheduling and technician-availability data on top of the same model. Hatch's CEO framed the addition carefully: "great voice AI requires more than a great voice model," pointing at the operational data behind it rather than the model alone. Neither company has disclosed commercial terms for the switch, and OpenAI's $0.05-per-minute API rate applies to the voice layer only -- tool calls and backend model usage bill separately.