- GPT-Live replaces the traditional turn-based voice pipeline with a full-duplex model that continuously listens and decides in real time whether to speak, pause, interrupt, or use tools.
- The key architectural pattern is delegated reasoning: a fast conversational model maintains the interaction while offloading complex reasoning or tool work to a larger background model (GPT-5.5).
- This split prevents higher reasoning effort from freezing the conversation, enabling more natural pacing with backchannels, interruptions, and mid-course corrections.
- Benchmarks like τ-Voice show current full-duplex agents still lose significant accuracy in real audio (about 30–45% of text-model performance) and often fail basic tasks like capturing names/emails or avoiding false claims.
- The main risk is the “seam” between models, where the front model may hallucinate to fill silence or mishandle the handoff when the background result arrives.
GPT-Live listens and speaks at the same time, deciding many times a second whether to talk, pause, interrupt, or reach for a tool. That replaces the turn-based pipeline, voice in, transcribe, think, speak back, that every voice agent was built on.
The pattern worth stealing is delegated reasoning. The fast conversational model holds the floor while it hands hard questions to a bigger model, GPT-5.5, running in the background. Turning reasoning effort up no longer freezes the call.
The new τ-Voice benchmark says the unglamorous parts still fail. Full-duplex agents keep only 30 to 45 percent of text-model accuracy under real audio, and they blow the basics: capturing a name, capturing an email, and not claiming they did something they never did.
On July 8 OpenAI shipped GPT-Live to every ChatGPT user, and the part that matters is not that the voice sounds better. It is that the model stopped taking turns. GPT-Live processes audio inside one model and decides, many times per second, whether to speak, stay quiet, interrupt, or call a tool. It drops in short cues like "mhmm" while you are still talking. You can cut it off mid-sentence, pause to think, or ask it to slow down, and it adjusts without restarting the exchange. More than 150 million people use ChatGPT Voice every week, which makes this the largest deployment of full-duplex speech anyone has run.
The pipeline everyone built is now the bottleneck
For three years a voice agent was a relay. Detect that the user stopped talking, transcribe the audio, send the text to a model, wait for the full response, run it through text to speech, play it back. Each stage added latency, and the whole thing moved in one direction at a time. Barge-in, the ability to interrupt, was bolted on as a special case, and it usually meant killing the current response and starting from scratch.
Human conversation does not work like that. People start responding within 100 to 300 milliseconds of a turn ending, and they overlap constantly, with backchannels, false starts, and corrections. A relay cannot hit that budget or that texture. Full-duplex collapses the relay into a single model that is always listening and always ready to speak, which is what makes the back-and-forth feel like a conversation instead of a walkie-talkie exchange.
Delegated reasoning is the pattern to steal
A model fast enough to hold a live conversation is not the model you want doing a multi-step lookup or a careful piece of reasoning. GPT-Live's answer is to split the job. The conversational model keeps talking, and when a request needs web search, deeper reasoning, or real agent work, it hands that task to a frontier model in the background. At launch that background model is GPT-5.5. The result comes back into the conversation when it is ready, and the fast model fills the gap in the meantime.
The payoff is concrete. Turning reasoning effort up to medium or high no longer stalls the call, because the heavy thinking happens off to the side while the light model keeps the line warm. This pattern generalizes past voice. Put a cheap, fast model in front of the user for presence and control, and delegate the expensive work to a slower model that does not have to answer in 200 milliseconds.
The seam is where it breaks
Splitting the work creates a seam, and the seam is the hard part. While the background model thinks, the front model has to say something, and the failure mode shows up the moment you picture it. The fast model, under pressure to fill silence, invents an answer the slow model has not returned yet. Or the slow model comes back and the fast model fumbles the handoff, talking over the result or dropping it on the floor.
This is not hypothetical. A new benchmark called τ-Voice, which extends the τ2-bench customer service suite to full-duplex audio, targets exactly these moments, the ones where a caller asks something research-heavy and then either waits or interrupts. The whole point of the test is whether the fast model stays honest while the heavy model works, and whether it recovers cleanly when the answer lands.
The boring parts still fail
The τ-Voice numbers are a useful cold shower. A text model with reasoning clears about 85 percent of these grounded tasks. The same tasks handed to voice agents drop to 31 to 51 percent under clean audio, and 26 to 38 percent once you add background noise, accents, and channel degradation. That is 30 to 45 percent of text-model capability surviving the move to voice.
The failures are not exotic. The single biggest one is authentication. Agents cannot reliably transcribe a name or an email, even when the caller spells it out letter by letter, and that one miss blocks everything downstream. Close behind is the hallucinated tool call, the agent announcing "I have updated your address" without ever making the call. Accents alone cost an average of 10 points, with a wide spread between providers, from a 1 point drop to an 18 point drop. When the researchers hand-checked 91 failed runs, 79 to 90 percent of the failures came from the agent, not from the test harness.
What to build for
If you are shipping a voice agent, the architecture lesson and the reality check point the same direction. Full-duplex plus delegated reasoning is the right shape, and it is worth adopting even before OpenAI opens the GPT-Live API, which is not available yet. The accuracy lives in the seam and in the plumbing, not in the voice.
Test the interruption paths, not just the happy path. Test what the front model says during the two seconds the background model is thinking, and make sure it does not promise a result it does not have. Nail data capture before anything else, because a voice agent that cannot write down an email address correctly will not survive contact with a real customer, no matter how natural it sounds while failing. The conversation got easier. The job it is having the conversation about did not.
If you’re building or choosing voice agents, prioritize systems that can stay responsive while delegating heavy work to a stronger model, but design explicitly for the handoff: make “waiting” behaviors safe, confirm key details like names and emails, and avoid stating outcomes until tools or the background model have actually returned results. When evaluating vendors, test real audio scenarios with interruptions and research-heavy questions, because that’s where full-duplex performance and honesty tend to break. Keep an eye on benchmarks like τ-Voice as a reality check for whether conversational smoothness is coming at the cost of correctness.