What Shipped
OpenAI has released its newest speech model, GPT-Live-1, to API developers. This is not a routine version bump. The model folds speech understanding and speech generation into a single native system, removing the traditional three stage chain of transcription into language model inference into synthesis that most real time voice products still run on.
For teams building voice agents, phone support bots or language practice apps, the practical payoff is lower latency, cleaner handling of interruptions, and a price point that changes the economics of the whole category: the front end voice layer bills at five cents per minute.
Why Full Duplex Is the Dividing Line
For the past two years almost all so called real time voice AI has been pseudo real time. A user finishes a sentence, the system waits for voice activity detection to declare the turn over, then runs transcription, reasoning and synthesis in sequence. Response delays commonly land between eight hundred milliseconds and two seconds. Worse, these systems handle interruptions badly. If a user speaks over the assistant, the interjection is either ignored or triggers a messy reset.
GPT-Live-1 works differently. It listens and speaks at the same time, managing interruptions, pauses and background noise dynamically during the exchange. OpenAI describes this as trained behavior rather than a rules engine bolted on afterward, with hearing and speaking optimized as one end to end task.
The gap shows up in production metrics. Early evaluation data from the language learning platform Speak indicates that misfires during learner thinking pauses dropped by nearly eighty percent compared with the turn based systems used before. In other words the model distinguishes someone pausing to think from someone who has finished talking, which is precisely the failure mode that made earlier voice bots feel rude.
From Demo to Production in Phone Workflows
If full duplex solves talking like a person, native telephone support solves working like one. OpenAI states that GPT-Live-1 already powers full duplex voice agents for use cases such as restaurant reservations and customer service. Developers can connect it to tools like Codex, or build on the OpenAI Presence framework to create workflows that query company systems, execute approved actions and escalate to a human when needed.
A technical detail worth attention: GPT-Live-1 supports handing complex reasoning and tool calls off to a backend text model. The front end voice model concentrates on fast perception and natural delivery while deeper thinking is delegated to a stronger text system such as GPT-6 Astra. That split architecture preserves conversational fluency without giving up the ability to resolve complicated business logic.
The benchmark numbers reflect the design. On the Tau3 end to end voice agent test, GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first. On Full Duplex Bench it scored thirty percentage points above the previous GPT-Realtime-2.1 model.
Engineering simplification shows up too. Tony Stoyanov, co-founder and CTO of a healthcare services platform, said his team cut eighty percent of their code, deleting roughly twenty three thousand lines of redundant logic after adopting the model. When the voice layer itself is capable enough, developers stop needing to stack state machines and exception handlers around it.
Reading the Price Correctly
Pricing often decides whether a technology escapes the experiment phase. The voice layer of GPT-Live-1 costs five cents per minute, which works out to about three dollars per hour of conversation. Important framing: that covers the speech front end only, while backend models and agent tool calls bill separately. Even so, the number opens a viability window for high volume voice work.
Run a common support scenario. A five minute phone call costs twenty five cents in pure voice layer charges. Human agent handling for a comparable ticket typically runs from a couple of dollars to well above that, which means the cost structure now supports replacing a meaningful slice of human seat work.
OpenAI also widened its voice lineup, adding twelve new voices named Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder, with additional voices and language support planned over the coming months. For products that need a distinctive brand voice, that selection matters.
What Still Remains Hard
The launch moves real time voice AI out of the technology showcase and into infrastructure. Full duplex operation, low latency, telephone native support and workable pricing combine for the first time in a way that lets developers ship a genuine voice product rather than another impressive demo that never reaches production.
The open problems have not vanished. Accent coverage, mixed dialect handling, robustness in extreme noise, and compliance and security requirements for enterprise deployment still need verification case by case in real business conditions. The tooling, though, has gotten materially sharper, and the API is already open.
