You pause to remember a booking number, and the assistant starts answering before you finish. You correct a delivery date, but it keeps talking. These small failures can make a voice assistant exhausting to use, even when it understands the subject. GPT-Live-1 tackles the timing of conversation: when to listen, when to speak, and when to give someone another moment.
OpenAI introduced GPT-Live in ChatGPT on July 8, 2026, then released GPT-Live-1 in the API on September 10. Its defining feature is full-duplex audio, which lets it listen and speak simultaneously. It can also hand demanding tasks to a backend model while keeping the conversation going. OpenAI’s GPT-Live introduction, API release announcement.
How GPT-Live-1 Handles a Conversation
To understand the change, consider how a conventional voice assistant works. Speech recognition turns your words into text, a language model prepares an answer, and speech synthesis reads it aloud. Each handoff takes time. Native speech models brought listening and speaking into one model, but earlier systems still relied on detecting when the user had finished a turn before responding. A pause could become an accidental invitation to speak. OpenAI’s voice architecture overview.
GPT-Live processes incoming audio continuously while generating speech. That gives it opportunities to acknowledge what someone is saying, yield to an interruption, or remain quiet. The benefit is easy to picture in language practice: a learner searching for a word needs time to finish the sentence. An immediate answer may be fluent and correct, yet undermine the exercise. How GPT-Live handles conversation.
The other significant change happens behind the conversation. GPT-Live-1 handles the spoken exchange while a separate backend performs delegated reasoning and uses tools. Developers can choose that backend independently, including connecting their own model or service. OpenAI contrasts this with Realtime, where one model handles speech, reasoning, and tool selection. For teams with an existing assistant, this separation offers a way to add voice while retaining the system that already handles their business rules. GPT-Live API guide.
Where More Natural Voice Interaction Helps
Language tutoring provides an early example of the benefits. In early evaluations reported in OpenAI’s API announcement, Speak found that GPT-Live-1 reduced interruptions during learners’ thinking pauses by almost 80% compared with its previous turn-based systems. That customer-specific result points to a useful measure of a tutor: whether it leaves the learner enough room to speak. Speak’s evaluation.
Customer support offers another use for this architecture. Imagine asking to change a delivery, then remembering that the replacement must arrive before Friday. The conversation needs to accommodate the new detail while the order system checks available options. A responsive voice interface helps collect the request; the connected system still has to find an option that satisfies it. The same division of work applies to reservations and appointment scheduling.
It could also make voice more useful for working with an agent. Someone might explain a software problem, ask for an investigation, and add a clue while the agent works. OpenAI’s architecture keeps these background tasks separate from the live audio stream. That preserves the ability to converse, although a slow tool or reasoning process still delays the actual answer. OpenAI explicitly acknowledges that the voice layer cannot hide an arbitrarily slow backend. The engineering behind GPT-Live.
How GPT-Live-1 Compares With Other Voice Models
The performance claims need to be read with this architecture in mind. OpenAI reports a 30-percentage-point improvement over GPT-Realtime-2.1 on Full Duplex Bench. It also reports a first-place Tau3 result with GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort. The latter measures the combined system, including its backend; it is not a standalone reasoning score for the voice model. OpenAI’s evaluation summary.
The alternatives share several capabilities, but organize the work differently. The comparison below focuses on documented behavior and integration choices.
Model | Conversation and task handling | What matters in practice |
GPT-Live-1 | Listens while speaking and delegates deeper work to a separately chosen backend. | Existing backend services can remain in place when adding voice. OpenAI |
OpenAI Realtime models | One model handles speech, reasoning, and tool selection. | These responsibilities sit together rather than being split across a voice model and a reasoning backend. OpenAI |
Gemini 3.1 Flash Live Preview | Native audio, configurable thinking levels, and voice activity detection for interruptions. | Google documents cancellation of ongoing generation when an interruption is detected. Google |
Amazon Nova 2 Sonic | Adjustable pause sensitivity and asynchronous tool calling. | Bedrock availability and Amazon Connect integration may suit existing AWS deployments. AWS |
Interruption handling alone is therefore a weak reason to choose one model. The more useful questions are how well it handles pauses and corrections, and how readily it connects to the systems that complete a request. GPT-Live-1 offers flexibility in choosing a backend; Nova’s integration with Amazon Connect may carry more weight for a team already using that service. Feature lists establish what is available, while comparisons of conversational quality require testing under the same conditions.
Costs and Practical Limits
At launch, GPT-Live-1 costs $0.05 per minute for the voice layer. Sessions are billed by duration, per second, with backend models and tools charged separately. At that rate, 1,000 session minutes cost $50 before those additional charges. Interruptions also have a practical limit: stopping the assistant’s speech does not automatically cancel its background work. If a caller changes their mind about a reservation, the application must handle the pending booking too. Launch pricing, Billing and application responsibilities.
GPT-Live-1 has its clearest appeal where a conversation is likely to change direction: a learner hesitates, a customer revises a request, or a colleague adds context to a task already underway. Its value will depend on whether those exchanges become easier without losing track of the work. The booking number at the start of the call still needs to be correct. The assistant should simply let you finish saying it.