OpenAI Launches GPT-Live-1 for Voice Reservations, Frontend Voice Layer Changed, Backend Still Uses Yelp Merchant Data
OpenAI developer account demonstrates GPT-Live-1: during calls, guests can interrupt, add details, or change direction while listening and speaking. It showcases how Yelp uses it to handle restaurant reservations. The model is open to API on the same day, serving as a developer version of ChatGPT's full-duplex voice, with the frontend voice layer priced at $0.05 per minute, allowing heavier reasoning and actions to be handled by paired models and tools.
Yelp connects GPT-Live-1 to two sets of merchant tools. Yelp Host is aimed at restaurant calls, while Hatch is for service providers like housekeeping. The official statement claims both are among the first products to deploy this model, using a full-duplex architecture to handle pauses, interruptions, and overlapping speech. Since its launch in October 2025, Yelp Host has processed over 1 million restaurant calls; Hatch manages tens of millions of leads across voice, SMS, email, and web, with client-side business value written as tens of billions annually. The technical director stated that the rounds and accuracy outperform the old voice architecture, with significant improvements in reservation and ordering connection rates, and callers' sentences are more complete.
The voice layer only changes the frontend. The backend still uses Yelp merchant data: business information, availability, ordering, daily specials, and seating arrangements. In July, Host integrated with OpenTable, allowing phone reservations, changes, and cancellations within the US and Canada network, with real-time updates without manual entry by staff. In August, Yelp reservations and waitlist entered ChatGPT conversations, allowing users to complete bookings without leaving the chat. Live-1 supplements the phone side, rather than creating another web button.
Reservations never go according to script: changes in party size, high chairs, noise, and interruptions from others. Full-duplex changes these from "the system doesn't understand" to a continuing conversation. The API is not bound to a single reasoning model; developers can connect lightweight backends for high concurrency scheduling and heavier models for complex customer complaints.
The market mechanism is the local living phone entry. The buy side consists of restaurants and service providers lacking front desk reception; the sell side is OpenAI selling the voice layer by the minute, while Yelp sells industry data based on connections and reservations. Funds flow from merchant subscriptions and per-call billing into the $0.05 per minute bill, then back to the platform from reservation commissions. The beneficiaries are Yelp, which holds both inventory and voice models; the pressured side is the keypad voice menu and systems that only do web reservations. The event-driven aspect is the API launch and customer demonstrations, not a new reservation network acquisition.
The phone is still there, but the script has been removed. A model that can be interrupted without dropping the call first takes the most expensive minutes from the front desk.
Source: Public Information
ABAB AI Insight
OpenAI has broken down the voice that can interrupt in ChatGPT into components sold by the minute. Yelp does not rebuild the reservation system, only changes the receiver. The 1 million calls are the corpus and processes accumulated from the old architecture, and Live-1 is responsible for making these processes sound less like a menu tree. At $0.05 per minute, voice becomes a utility cost rather than a project-based model. The OpenTable interface proves that inventory still resides in the old network, and the new model competes for connection rates, not for creating a new set of availability.
The capital path is data companies renting voice and model companies renting minutes. Yelp's moat is the primary data of restaurants and the 1 million connected calls, while OpenAI's moat is full-duplex and interchangeable backends. Hatch sells the same receiver to plumbers and painters, expanding minute consumption. Reservations in ChatGPT are text-based, while phone reservations use Live-1, with both sides writing Yelp as the execution layer. A portion of local living advertising budgets is redirected: from clicks to answered calls.
The analogy is cloud call centers changing engines, and navigation software integrating ride-hailing without building their own fleets. The industry phase is controlling the dialogue entry. Whoever becomes the default receiver first will charge between employee wages and missed orders. The elimination of keypad IVR does not require new regulations, only that callers are willing to speak complete sentences.
The structural change is technological substitution. The mechanism is full-duplex canceling the machine politeness of "waiting for the other party to finish speaking"; phone labor shifts from staffing to a per-minute model. The pricing power of the reservation system remains with the availability inventory, while the connection power shifts to the voice layer suppliers. Scripted customer service loses naturalness first in the face of interruptible models, then loses connection rates.
ABAB News · Cognitive Laws
A model that can be interrupted without dropping the call first replaces the most expensive minutes at the front desk.
Inventory can still reside in the old network, but the receiver has changed.
After voice is priced by the minute, connection rates become a purchasable commodity.