OpenAI opens up GPT-Live-1's full-duplex voice model via API

OpenAI has taken its real-time voice model out of the ChatGPT app and put it into developers' hands. GPT-Live-1, the full-duplex speech model that debuted in July as a ChatGPT feature, is now accessible via API, meaning anyone building voice-enabled software can plug it in directly rather than waiting for OpenAI to ship it in a consumer product first.

The headline feature is "full-duplex" operation, which sounds technical but solves a very human problem. Most voice AI systems to date have worked in half-duplex mode: the system listens, waits for you to finish, then responds. It's the conversational equivalent of a walkie-talkie, and it breaks down the moment a real conversation happens, because people interrupt, backtrack, and talk over each other constantly. GPT-Live-1 can listen and generate speech at the same time, which lets it handle interruptions and overlapping speech in a way that feels less like talking to a phone tree and more like talking to a person.

That matters a lot for anyone building voice agents for customer service, tutoring, or booking systems, where stilted turn-taking is often the single biggest tell that you're talking to a bot.

Architecture and benchmark numbers

OpenAI has designed GPT-Live-1 to be a front-end conversational layer that sits on top of a separate backend model doing the actual thinking. The voice model handles the flow of spoken dialogue, while something like GPT-6 Astra, or a cheaper model for simpler tasks, handles lookups, tool calls, and task execution behind the scenes. That split lets developers tune cost and capability independently: you don't need your priciest model just to manage natural pauses and interruptions.

This is a generational jump from OpenAI's original Realtime API, which launched back in 2024 and set the baseline for low-latency voice interaction. The company's numbers back up the leap:

• On Tau3, a benchmark that scores spoken customer-service tasks across airline, retail, and telecom scenarios, GPT-Live-1 hits 86.2 percent.
• GPT-Realtime-2.1 scores 45.7 percent on the same benchmark.
• The older GPT-Realtime-2 trails at 42.4 percent.
• On OpenAI's internal Full Duplex Bench, GPT-Live-1 performs 30 percent better than GPT-Realtime-2.1, with improvements also showing up in turn-taking latency and general interactive behavior.

Those are OpenAI's own evaluations, worth the usual grain of salt reserved for vendor benchmarks, but the gap between 86 percent and roughly 44 percent on Tau3 is large enough to suggest a real architectural shift, not just incremental tuning.

Language-learning app Speak has already run GPT-Live-1 through its paces as an early tester. According to OpenAI, Speak found that the model's smooth interruption handling gave learners more time to think before the tutor jumped in, cutting interruptions by almost 80 percent compared to the company's previous turn-based system. For a language tutor, that's not a cosmetic improvement: awkward, premature interruptions are exactly what makes people clam up when practicing a new language out loud.

Yelp is the other named early adopter, running GPT-Live-1 behind its Yelp Host restaurant reservation service. Akhil Kuduvalli Ramesh, Yelp's chief product officer, said in a statement that "with GPT-Live-1, every call is more conversational and responsive," adding that Yelp Host now "gives restaurants an advanced voice AI that delivers exceptional guest experiences, captures revenue opportunities they may have otherwise missed, and helps staff focus on serving guests instead of answering the phone."

Worth noting: Yelp's statement is entirely framed around business benefits, not customer sentiment. There's no data on whether people calling a restaurant actually prefer a more free-flowing AI voice on the other end of the line, or whether they simply tolerate it because it beats sitting on hold. That's a gap developers building on this API should probably fill with their own user research rather than assume away.

Pricing is straightforward on paper: $0.05 per minute of GPT-Live-1 usage, stacked on top of whatever backend model is doing the reasoning work. That's an important detail for anyone estimating costs, since a chatty voice interface running on top of a frontier reasoning model could add up fast for high-volume applications like call centers. Custom voices aren't self-serve, they require going through OpenAI's sales team directly. The model is also baked into OpenAI Presence, the company's enterprise agent platform, for businesses that want the voice layer bundled with broader agent tooling rather than wired up piecemeal via API.

Key Takeaways

GPT-Live-1's move from a ChatGPT feature to a general-purpose API is the more consequential story here, even more than the benchmark gains. It puts full-duplex, interruption-tolerant voice interaction into the hands of any developer building a conversational product, not just OpenAI's own app.

For teams building voice agents, the architecture is the thing to pay attention to: a fast, natural-sounding conversational layer decoupled from a swappable backend reasoning model. That separation of concerns, speech handling versus task logic, is likely to become a standard pattern across the industry, not just an OpenAI quirk.

The benchmark improvements are real on paper, jumping from roughly 42-46 percent to 86 percent on Tau3 is not a rounding error. But the early evidence so far comes entirely from OpenAI's own tests and two customer case studies it chose to publicize. Anyone evaluating this for production use should expect to run their own latency and interruption-handling tests against real user behavior, especially before committing to per-minute pricing that scales with call volume.