Every AI phone agent you've ever called has the same tell. You talk, it waits. It talks, you wait. Today OpenAI released the model that ends that. The same day, it went live on WindowEdge, making us the first home improvement platform running it.
Here's the dirty secret of voice AI: most of it is a relay race.
Your words get turned into text. The text gets handed to a language model. The model's answer gets handed to a voice engine. Three runners. Two baton passes. Every single turn.
Every handoff costs time. Every handoff drops something: the hesitation in a homeowner's voice, the "mm-hm" that means keep going, the moment they start talking over the agent because it got the address wrong.
That's why AI calls feel like walkie-talkies. Over.
What OpenAI Actually Shipped
GPT-Live-1 isn't a faster relay. It's no relay at all.
One model listens and speaks at the same time. It hears you while it's talking. It knows the difference between an interruption and an "uh-huh." It can stop mid-sentence, take the correction, and keep moving.
Engineers call it full-duplex. Homeowners call it "wait, that felt like a person."
OpenAI's own numbers against its previous best real-time voice model:
| GPT- |
GPT-Live-1 | |
|---|---|---|
| Turn-taking latency | 1.41 sec | 0.798 sec |
| Tool-calling success | 60% | 87% |
| Interactivity (Full Duplex Bench) | 45.4% | 80.1% |
Source: OpenAI, GPT-Live-1 launch, September 10, 2026.
Circle that tool-calling number. Tools are how an agent checks availability, validates an address, and actually books the appointment. That's the whole job.
Why This Matters on a Window & Door Call
Real homeowner calls are messy. Nobody reads from a script.
"We need three double-hungs in the... actually, it's the back slider that's the real problem."
"Tuesday works. Wait, no. We've got the kids' game Tuesday."
"Hang on, the dog's going nuts. Okay. Sorry. What was your question?"
A relay-race agent handles that like a GPS recalculating. It finishes its sentence, processes the old answer, and confirms the wrong day.
A full-duplex agent hears "wait, no" while it's still talking. It stops. It fixes it.
| Relay-race voice AI | GPT-Live-1 on WindowEdge |
|---|---|
| Talks over the homeowner | Yields the moment they cut in |
| Dead air while it "thinks" | Keeps the conversation moving while work runs in the background |
| Corrections get lost mid-booking | Corrections reach the backend before anything is confirmed |
| Every turn waits on three systems | One model hears and speaks at once |
Delivering the Cutting Edge, Every Day
Our dealers count on us to keep them on the best technology available. We try to earn that trust every day.
So when something as meaningful as GPT-Live-1 comes along, our instinct is to roll up our sleeves and start building with it right away, rather than add it to a roadmap.
The same day OpenAI opened the API, our engineering team built a GPT-Live-1 voice runtime into the WindowEdge platform and put a production dealer agent on it. The complete playbook. The caller context. Every tool it uses to do the job.
Here's the architecture:
The voice: GPT-Live-1. It owns the conversation. Listening, talking, interruptions, pacing.
The brain: GPT-5.6 class reasoning models, loaded with the dealer's full playbook. Every business decision is delegated behind the scenes: service area and ZIP checks, availability, booking, text confirmations, transfers to your team.
The gate: WindowEdge. Every incoming call is signed, verified, and matched to the right dealer, the right agent, and the right phone number before the agent says a word.
The rule: the voice never makes things up. No invented availability. No invented quotes. No invented confirmations. If it didn't come back from your real calendar and your real rules, the agent doesn't say it.
And it's built to know who's calling. A returning homeowner shouldn't have to spell their name for the third time.
And every call lands in WindowEdge with the outcome, a plain-English summary, the full transcript, and the recording, ready for your team to review.
Tested Every Week, on Real Phone Lines
We don't pick a model, sign a contract, and ride it for three years. We run a standing bake-off.
Every week, we test the latest voice AI models: speech-to-text, text-to-speech, and large language models from OpenAI, ElevenLabs, Deepgram, Cartesia, LiveKit, and more. Same dealer scripts. Real phone calls. We listen, and we score them.
And we learned something most vendors never find out: a model that dazzles in a browser demo can fall apart on a real phone line. One engine that sounded incredible on a laptop came through an actual cell call sounding like a ransom video. You'd never catch that on a spec sheet. You only catch it if you pick up the phone.
That's the gauntlet every model runs before it earns a dealer's phone line. Real calls. Reviewed transcripts. Every booking checked end to end.
We move at frontier speed. We don't experiment on your customers.
What This Means for Your Dealership
- You don't chase models. We do. Our team evaluates major releases the week they ship. This time, the day they shipped.
- Your playbook carries forward. Service areas, calendars, booking rules, and hand-offs stay with your agent when the engine underneath gets better.
- Your customers only hear the winner. Nothing reaches your phone lines until it proves itself on real calls.
Your competitors bought a voice bot in 2024.
It still sounds like 2024.
See It in Action
GPT-Live-1 went public today. It's already live on WindowEdge.
Schedule a Demo and hear what a home improvement call sounds like when the AI finally stops talking over your customers.
The frontier moved today. So did we.
