Summary
- OpenAI has released GPT-Live-1 in the API. It handles listening and speaking inside a single model, replacing the front end of the old three-part stack that turned speech into text and text back into speech.
- On Full Duplex Bench, which measures live back-and-forth, the model gains 30 percentage points over its predecessor, and paired with GPT-6 Astra it ranks first on Tau3, the company said.
- The language-learning app Speak reported that interruptions fell by almost 80 percent versus its previous system in early evaluations. The voice front end costs five cents a minute, with backend reasoning and tool use billed separately.
OpenAI has opened its voice model GPT-Live-1 to developers through the API. The model had been available only inside ChatGPT, and the point of it is that it listens and speaks at the same time. In its announcement the company wrote that "GPT-Live-1 handles listening and speaking in a single model, simplifying the voice layer." The front end carries the conversation; anything that needs real judgment is passed to a second model sitting behind it.
Until now, voice agents have been assembled from three pieces: a module that turns speech into text, a model that thinks, and a module that turns text back into speech. Every handoff, the company explained, adds latency and creates more chances to lose timing, context, or the natural rhythm of a conversation. It is interpretation run as a relay race, and the runners lose a beat each time the baton changes hands.
The structural change shows up in the scores. On Full Duplex Bench, which measures live give-and-take, the model improves by 30 percentage points over GPT-Realtime-2.1, with the largest gains in turn-taking latency and interactive behavior, the company said. Paired with GPT-6 Astra at medium reasoning effort, it ranks first on Tau3, which grades voice-agent intelligence on end-to-end tasks.
The company also published what the evaluations cover. Spoken customer-service tasks in airline, retail, and telecom domains were scored for task success, and a banking suite counted how many of 97 tasks combining knowledge retrieval with account tools were completed. Separate measures cover how the model handles pauses, turn-taking, interruptions, and backchannels, how quickly it starts replying once a user stops talking, and whether it still calls the right tools when a request is full of hesitations and self-corrections.
One deployment came with a number attached. The language-learning app Speak found in early evaluations that GPT-Live-1 gave learners more time to think and cut interruptions by almost 80 percent compared with the previous turn-based system, the company said. Waiting for someone who speaks slowly turns out to weigh as much as a feature in a voice product.
METAL watched the full 76-second public demo, which sets out to stress exactly that behavior. A presenter talks over the model mid-sentence, then pipes in background noise and asks whether he can still be heard. The model answers that it is coming through clearly and can follow his voice even with noise around, then picks up where it left off. In the same clip it hands actions to a robot and a display while the conversation keeps running.
Folding the voice front end into one model also cuts the code developers used to write themselves. Previously a developer had to decide, step by step, what should happen when a user interrupted, paused, or changed direction. The company wrote that developers were the ones left coordinating those stages, and said the front end now absorbs interruptions and acknowledgements on the spot while passing only the deeper judgment backward. Work proceeds in the background without the conversation stopping.
What sits behind the voice is the developer's choice. The company offered pairings as examples: a model like Luna for high-volume work such as rescheduling or order updates, and Astra for complicated customer problems that need reasoning. Keeping the same voice front end and swapping only the back end lets teams tune depth, speed, and cost task by task.
What the front end passes along is spelled out too. GPT-Live-1 returns both the transcript and the response text, understands speech mixing digits and letters, and supports keyword biasing. Although it is not a turn-based model, it still detects turn boundaries natively, so teams that built around explicit turns do not have to tear up their architecture.
The price covers only the front end. The company set the voice layer at five cents a minute, and the demo notes immediately that this is the front-end price, with backend inference and tool services billed separately. What a call actually costs therefore depends on what is bolted on behind it, so budgeting from the per-minute rate alone will be wrong.
Voice choice is wider but still growing. The company wrote that it has expanded from a small set of real-time voices to a broader selection across accents, dialects, and languages, while adding that it will keep expanding voice options and language availability over the coming months. Custom voices are for now limited to customers who clear an eligibility check through a sales contact.
There is also an enterprise route. The company introduced OpenAI Presence, which uses the model to power real-time voice interactions, describing agents that answer questions, resolve issues, use company systems, take approved actions, and escalate to a person when needed. The ambition reaches past selling a voice front end to selling the workflow around it.
The delegation pattern lines up with the direction OpenAI has been pushing. METAL reported on the data agent OpenAI put into ChatGPT Work to answer questions from a company's own files, where the front end took the request and the actual work happened in tools behind it. Voice has now walked in through the last remaining door of that structure.
In short, OpenAI has replaced one of the three pieces developers used to stitch together with a model of its own. The whole design is a division of labor in which the front end absorbs interruptions and noise while judgment moves backward. Two things are worth watching. Whether the rhythm of the conversation survives a change of backend model, and what standard companies settle on for pricing the backend work that stacks on top of the per-minute rate.
Comments