Every voice agent you have ever hung up on failed at the same moment. You tried to interrupt it.
You said no, that is not what I meant, and the system kept talking, because it was waiting for a detector to decide your turn had ended. The vocabulary was fine. The conversation was not.
OpenAI published an engineering write-up on GPT-Live, its third-generation voice system, and the architectural decision at the center of it is the one that matters commercially: the turn detector is gone from the audio path.
Full duplex is a behavior change, not a speed bump
The model is full-duplex. It listens and speaks at the same time, treating a conversation as one continuous audio stream rather than a sequence of discrete exchanges.
That single change removes the walkie-talkie quality that made every previous generation feel like a machine. There is no moment where the system stops hearing you in order to reply.
The supporting numbers are real engineering rather than marketing. Moving the media frontend and inference logic to Go from Python asyncio improved frame delivery enough that p95 latency now matches what p50 used to be. Median time to the first audio chunk sits in the 300 to 600 millisecond band.
There is also a routing decision worth noting. When a request needs deeper reasoning or a tool call, the system can consult a frontier model without stalling the conversation, so the fast path stays fast and the slow path stays available.
The part that changes your P&L
I build these systems, so let me be direct about where the commercial value actually sits.
The reason voice deployments stall is almost never the script. It is containment: the percentage of calls the agent finishes without a human. And containment dies at interruption.
A customer corrects the agent mid-sentence. The agent does not register the correction, continues down its branch, and now the customer is annoyed and the transfer is guaranteed. Every one of those calls costs you twice, once for the automation and once for the human who cleans it up.
Remove the turn detector and that failure disappears from the architecture rather than being patched at the prompt level. You stop writing elaborate instructions telling a model how to behave when interrupted, because interruption is no longer an edge case in the pipeline.
My working expectation is that this moves containment materially on exactly the call types that were previously borderline: order changes, appointment reshuffling, anything where the caller revises themselves halfway through a sentence.
There is a second economic effect that gets less attention. Turn-based systems pay a latency tax on every exchange, and long calls accumulate that tax until the caller gives up. Cutting the pause does not just improve tone, it shortens handle time, and handle time is the number your operations lead is actually judged on.
Worth being precise about where this does not help. A full-duplex model does not fix a bad knowledge base, a missing CRM integration, or a policy nobody can explain. It removes one specific class of friction, and it removes it well.
What to do about it this quarter
If you already run a voice agent, your evaluation set is probably wrong. Most teams test clean, cooperative callers reading well-formed requests.
Rebuild the test set around messy behavior. Interrupt the agent at three seconds. Change your mind twice in one sentence. Talk over the confirmation. Go silent for four seconds and then continue. Those are the calls that decide whether the system survives contact with real customers.
Then measure the right thing. Not satisfaction scores, which callers give generously, but resolution without transfer, and repeat contacts within seven days. Those two numbers tell you whether the agent solved anything or simply postponed the human.
If you have not deployed one yet, the honest advice is that the architecture just crossed a threshold, and the constraint has shifted back to where it usually sits: your data and your process.
Start with one call type rather than a department. Pick the highest-volume, lowest-variance reason people phone you, automate that alone, and keep everything else routed to humans until the first one holds for a month.
An agent that sounds human and still cannot see the order status is a faster way to disappoint somebody. The integration work has not gotten easier just because the voice did, which is still why most AI pilots die in the middle.
At difrnt.ai we build these end to end, and the pattern is consistent. The model layer is now the least interesting part of the project. The interesting part is what the agent is allowed to know and allowed to do, and not letting one vendor own your agents while you find out.
Six months of engineering removed the pause. It did not remove the requirement to have something worth saying.