Yelp Integrates OpenAI's GPT-Live-1 Voice Model to Automate Local Business Calls

Local restaurant operators and home-service businesses can now deploy OpenAI's newest full-duplex voice model to answer their phones, following Yelp's integration of GPT-Live-1 into its commercial software suite. The update swaps the front-end voice layer of Yelp Host and Hatch for a system that listens and speaks simultaneously, allowing callers to interrupt, pause, or switch languages mid-sentence without breaking the AI's conversational thread. By separating the live voice interaction from the backend business logic, the system holds the conversation while simultaneously checking a reservation book or pulling a service calendar.[1][4]

The technical shift addresses the primary bottleneck in automated customer service. Traditional voice agents chain together speech-to-text transcription, a text-based reasoning model, and text-to-speech synthesis. Each handoff adds latency and forces strict turn-taking, resulting in the robotic pauses that frustrate callers. GPT-Live-1 collapses this stack into a single model that processes incoming and outgoing audio natively, cutting turn-taking latency to 0.798 seconds—down from 1.41 seconds in the previous GPT-Realtime-2.1 architecture.[2][3]

For local businesses, the practical stakes are tied directly to lead conversion and labor costs. Yelp Host, which launched in October 2025, has already processed more than one million calls for restaurants. With the GPT-Live-1 upgrade, Yelp reports a measurable drop in call transfers to human staff. Yelp chief technology officer Alex Levy noted that the improved accuracy changes caller behavior. "Callers are also speaking fuller, more natural sentences, which tells us the experience on the other end of the phone feels genuinely different," Levy said.[1][3]

The integration extends to Hatch, a communication platform Yelp acquired in 2026 to serve home-service professionals like plumbers, electricians, and contractors. Hatch manages tens of millions of leads across multiple channels. When a homeowner calls, the voice agent diagnoses the issue, verifies the service area, and books the appointment directly into the calendar. "GPT-Live-1 transforms how the conversation feels, and Hatch provides the operational intelligence that turns those conversations into booked jobs," said Chris Bache, co-founder and chief executive officer at Hatch.[1]

Because the model processes audio natively rather than relying on text transcripts, it detects caller tone and adapts its pacing accordingly. Yelp layered its own proprietary business data and language enhancements on top of the OpenAI foundation. This allows the system to automatically detect and respond in multiple languages, preventing lost leads when a caller speaks a language the on-duty staff does not.[1]

The developer economics behind the deployment highlight OpenAI's strategy for enterprise adoption. OpenAI released GPT-Live-1 to its API on September 10, 2026, at a cost of $0.05 per minute, billed per second. However, this price covers only the front-end voice layer. Backend model usage and tool calls—the operations that actually check the database or process the booking—are billed separately.[2][3]

This decoupled architecture allows developers to optimize their infrastructure costs. A voice agent can use the lightweight GPT-Live-1 model to handle basic conversational back-and-forth, such as greetings and acknowledgments, and only wake up a heavier, more expensive reasoning model when a complex query demands it. By assigning the conversational rhythm natively to the API, developers no longer have to hard-code the complex turn-taking logic that previously made voice agents brittle.[2][3]

Performance benchmarks published by OpenAI indicate a significant jump in end-to-end task completion. When paired with GPT-6 Astra at a medium reasoning effort, GPT-Live-1 completed 83.6 percent of tasks on the Tau3 benchmark—which evaluates AI agents on airline, retail, and telecom support scenarios. That represents a near-doubling of the 45.7 percent success rate achieved by the previous GPT-Realtime-2.1 model. On the Full Duplex Bench, which measures interactive behavior and interruption handling, the new model scored 80.1 percent.[2][3]

The integration ties Yelp's platform more directly to the day-to-day operations of local businesses, moving beyond its historical reliance on advertising and text reviews. As generative AI search tools threaten traditional local discovery, embedding operational software like Host and Hatch provides Yelp with a stickier, subscription-based revenue stream. By automating the front desk, the company positions its software not just as a marketing expense, but as a direct offset to labor shortages in the hospitality and service sectors.[1][4]