OpenAI has made GPT-Live-1 available through its API, giving developers access to a voice model that can listen and speak simultaneously rather than waiting for a turn. The model debuted inside ChatGPT in July; the API release moves it into customer-facing products, priced at five cents per minute of use on top of whatever backend model handles the actual work.
Key takeaways
- GPT-Live-1 is full-duplex, meaning it generates speech and processes incoming audio at the same time instead of alternating like a walkie-talkie.
- It scores 86.2 percent on the Tau3 voice-agent benchmark, against 45.7 percent for GPT-Realtime-2.1 and 42.4 percent for GPT-Realtime-2.
- Usage costs $0.05 per minute plus backend model charges, and the model is also offered through OpenAI Presence, the company's enterprise agent platform.
Why full duplex changes the feel of a voice agent
Earlier voice systems are half-duplex: one side talks, the other listens, and the handoff is explicit. Human conversation does not work that way. People overlap, interject, trail off and get gently interrupted, and a system that cannot do any of that reads as stilted no matter how good its answers are.
Listening while speaking lets the model notice an interruption as it happens rather than after it finishes a sentence. OpenAI points to an early deployment by the language-learning company Speak, where the model reportedly cut interruptions by close to 80 percent compared with turn-based predecessors by giving learners longer to think before the tutor responded. The improvement there is restraint rather than speed — knowing when not to talk.
The benchmark gap is unusually wide for a single generation. Tau3 measures spoken customer-service tasks of the kind airlines, retailers and telecom operators run, and the jump from the low forties to 86.2 percent suggests the previous Realtime models were failing at something structural rather than falling marginally short. OpenAI separately reports roughly 30 percent better Full Duplex Bench performance than GPT-Realtime-2.1, plus gains in turn-taking latency.
How it fits into an application
GPT-Live-1 is not meant to work alone. It handles the conversation while a separate backend model does information lookup, tool calls and task management — GPT-6 Astra for demanding work, or a cheaper model where the task is routine. That split keeps the expensive reasoning off the per-minute voice meter and lets teams tune cost per conversation rather than accept one blended rate.
It also means the pricing needs reading carefully. Five cents a minute covers the voice layer only. A ten-minute support call costs fifty cents before the backend model, the agent logic or the telephony bill, so unit economics depend on which model sits behind the microphone. Custom voices are available but must be negotiated with OpenAI's sales team rather than configured.
Yelp is among the first named users, running the model in Yelp Host, its AI restaurant reservation service. The company's chief product officer, Akhil Kuduvalli Ramesh, said the technology makes calls more conversational and helps staff spend time with guests instead of the phone. As The Register observed, nobody has yet asked the diners on the other end how they feel about it.
What developers should test before committing
The benchmark improvements describe average behaviour, and voice deployments fail at the edges. Accents, background noise, poor phone lines and speakers who pause mid-sentence are where half-duplex systems traditionally break, and where a model that interrupts confidently can be worse than one that waits. Any evaluation worth running should include bad audio, not clean studio samples.
Disclosure is the second issue. A system fluid enough to pass for a person on a reservation line raises questions about whether callers are told, and several jurisdictions now require that they are. Teams shipping this into consumer voice flows should settle the policy before the launch, not after the first complaint.
The wider direction is clear enough. OpenAI is packaging capability as infrastructure others build on, the same pattern visible in its enterprise agent platform and in the ambient-listening assistants arriving elsewhere in the industry. Voice is becoming a layer rather than a feature.
FAQ
What does full-duplex mean for an AI voice model?
It means the model can produce speech and take in audio at the same time, rather than strictly alternating between listening and talking. That allows it to detect and respond to interruptions mid-sentence, which makes conversations feel closer to human exchanges than to radio handsets.
How much does GPT-Live-1 cost?
The voice model costs $0.05 per minute of usage. That is charged on top of the backend model handling lookups, tool use and task management, so total cost depends on whether a team pairs it with a capable model like GPT-6 Astra or a cheaper option for routine work.
How does GPT-Live-1 compare with OpenAI's Realtime API?
On the Tau3 voice-agent benchmark GPT-Live-1 scores 86.2 percent, compared with 45.7 percent for GPT-Realtime-2.1 and 42.4 percent for GPT-Realtime-2. OpenAI also reports about 30 percent better Full Duplex Bench results than GPT-Realtime-2.1, along with improved turn-taking latency.
