Yelp Host and Hatch's New AI Voice: What Actually Changed

Yelp's AI receptionist used to wait its turn to talk. As of September 10, it doesn't have to anymore. Yelp Host, the voice AI answering restaurant phones, and Hatch, the AI receptionist platform Yelp bought for home service businesses back

Yelp's AI receptionist used to wait its turn to talk. As of September 10, it doesn't have to anymore.

Yelp Host, the voice AI answering restaurant phones, and Hatch, the AI receptionist platform Yelp bought for home service businesses back in January, both just swapped their voice engine for OpenAI's new GPT-Live-1 model. Callers can now interrupt mid-sentence, change the subject, or yell an order across the kitchen, and the AI keeps up instead of freezing and restarting.

It's a real upgrade to a real problem. It's also a smaller upgrade than the announcement makes it sound, and knowing the difference matters if you're a restaurant or home service business deciding whether your Yelp listing's free-ish AI receptionist is finally good enough.

Yelp and Hatch Both Rebuilt Their Voice Layer on the Same Model

Yelp announced the integration on September 10, putting OpenAI's GPT-Live-1 behind both Yelp Host and Hatch as the new front-end voice layer. Yelp Host handles restaurant calls: reservations, wait times, and since a July update, takeout orders routed straight into the kitchen. Hatch, which Yelp paid roughly $270 million in cash plus $30 million in retention payments to acquire in January, does the same job for plumbers, HVAC companies, auto shops, and other businesses that dispatch a technician instead of seating a table.

Akhil Kuduvalli Ramesh, Yelp's Chief Product Officer, framed it as a race to be first: "We're excited to be among the first to integrate GPT-Live-1 into Yelp Host and provide the leading AI voice solution for restaurants." Hatch's co-founder and CEO, Chris Bache, made a similar case for service businesses.

Hatch is among the first platforms to integrate OpenAI's GPT-Live-1, and we're thrilled to deliver a new, more natural real-time voice AI experience for service businesses.

— Chris Bache, Hatch co-founder and CEO

Both products kept their existing backend, the part that actually knows your menu, your service area, or your technician's schedule. What changed is the layer that talks to the caller. OpenAI's own multimodal product lead, Teri Yu, described the release as infrastructure for exactly this kind of deployment: "We're bringing GPT-Live-1 to the API so developers can build more natural, powerful voice experiences by pairing it with their choice of models and tools." Yelp and Hatch are two of the first companies to actually ship it.

Yelp Host and Hatch aren't the only ones racing on this. Zoom, RingCentral, and Intermedia have all shipped some version of a built-in AI receptionist this year, and Podium and Birdeye have been iterating on their own voice AI agents through 2026. When every phone platform and local-business tool ships roughly the same feature within months of each other, the voice itself stops being the differentiator. It becomes table stakes, and the competition moves to whatever's underneath it.

Full-Duplex Voice Fixes the Thing That Made AI Calls Feel Fake

Most voice AI up to this point has been half-duplex: the AI listens, waits for silence, then responds. That's why older AI receptionists go quiet mid-interruption, talk over a caller who jumps in, or need a caller to repeat themselves after a false start. It's a small delay, but it's the exact tell that gives away you're talking to a bot.

GPT-Live-1 is a full-duplex model, meaning it can listen and generate speech at the same time, the way a real conversation actually works. Layered into Yelp Host and Hatch, that means a caller can interrupt to correct an order, change topics mid-sentence, or have a kid yelling in the background, and the AI adjusts instead of stalling. Yelp says the model also reads tone, so a frustrated or rushed caller gets handled differently than a relaxed one, and it can switch languages automatically without a menu prompt.

That volume is exactly why the timing makes sense. Yelp Host crossed 1 million calls in July, growing at 38% a month on average since launch. At that scale, every awkward pause is a caller who hangs up and calls a competitor instead.

A Better Voice Doesn't Make a Smarter Receptionist

Here's the part the press release doesn't dwell on, and worth saying plainly if you're evaluating this as a buyer instead of reading it as tech news.

GPT-Live-1 is OpenAI's mouth and ears. It is explicitly not the brain. OpenAI's own documentation is blunt about this: the model "delegates real reasoning to a separate backend model," meaning whatever business logic decides what the AI is allowed to promise, book, or say still lives in whatever Yelp and Hatch built themselves. A more natural voice sitting on top of a limited system is still a limited system, just a more convincing one.

We've made the same point about the gap between a built-in AI receptionist and a dedicated one: the voice was never the hard part. The hard part is whether the AI can reach your actual booking system, your actual technician calendar, your actual order rules, past whatever the platform decided to expose. GPT-Live-1 doesn't change that answer for Yelp Host or Hatch. It just makes the limitation harder to hear, because the call sounds confident right up until it can't do the thing you called for.

That's the whole argument for building an AI receptionist around your business specifically instead of subscribing to whichever generic version your listing platform ships this quarter. The voice model underneath it can be swapped out by any vendor at any time, ours included. What can't be swapped out cheaply is the work of wiring the AI into your calendar, your CRM, and your actual rules for what it's allowed to promise a caller.

The Bundled Price Hides What Voice AI Actually Costs to Run

Worth knowing if you're comparing a platform subscription to building something yourself: GPT-Live-1 isn't free to run, even at OpenAI's own list price. Sessions bill at five cents a minute for the voice layer alone, calculated per second, with the backend reasoning model and any tool calls charged separately on top. That's before you add the business logic, the integrations, and the actual engineering.

Yelp and Hatch are absorbing that infrastructure cost into a flat monthly subscription, which is a fine deal if the flat price is right for your call volume. It's the same tradeoff we walked through when comparing Retell, Vapi, and Bland as raw voice AI infrastructure instead of a finished product: usage-based platforms get cheaper per minute as volume scales, but someone still has to build and maintain the part that makes the AI useful for your specific business. Yelp and Hatch did that work for you. The cost of that convenience is that you're running their menu of features, not yours.

What to Check Before You Trust the Upgrade

None of this means skip the update or distrust it. A voice that doesn't stall on an interruption is a genuine improvement, and if you're already paying for Yelp Host or Hatch, you get it automatically with no action required.

What it does mean is judge the upgrade for what it is. Book a real test call and try to break it the way an actual customer would: interrupt it, change your order halfway through, ask something outside its script. Listen for whether it recovers gracefully now that it can, then check whether it actually completes the task you called about, a booking, a reschedule, a real answer, not just a smoother conversation on the way to a dead end.

Three questions worth answering before you decide the upgrade is enough:

Can it reschedule against your real availability, not just book into an open slot? Can it quote a price or a wait time that's actually current, not a number baked into its training? And when it hits the edge of what it knows, does it hand off to a human cleanly, or does it just keep talking?

If your calls are simple, hours, directions, a basic booking, this upgrade alone might be reason enough to stick with the bundled option. If your business runs on exceptions the way most HVAC and service calls do, a nicer voice on the same limited backend won't close that gap. That's a different build, not a software update.