OpenAI has released GPT-Live-1 through its API, giving developers access to the full-duplex voice technology already used in ChatGPT Voice. The model can listen while speaking, react to interruptions and pauses, and hand off more difficult reasoning or tool work to a separate backend model without ending the conversation.
The API launched on 10 September 2026 at $0.05 per minute for the front-end voice layer. That price does not include the reasoning model, tools or other services sitting behind the conversation.
Before comparing it with standard text-to-speech, it helps to separate the two jobs. DIY AI’s AI voice generator is designed to turn text into generated speech. GPT-Live-1 is designed for continuous two-way interaction, where silence, interruptions, timing and changing instructions are part of the input.
According to OpenAI’s GPT-Live-1 API launch, developers can also choose the backend model, tools, and agent harness behind the voice layer. That architectural choice may prove more important than another improvement in synthetic voice quality.
What OpenAI actually launched
| GPT-Live-1 feature | What it changes |
|---|---|
| Full-duplex audio | The model can listen and speak at the same time instead of forcing rigid user-agent turns. |
| Backend delegation | Reasoning and tool calls can be handed to models such as GPT-6 Astra, Luna or a third-party backend. |
| Promptable delivery | Developers can control tone, pace and conversational style. |
| Built-in transcription | GPT-Live-1 provides ASR transcripts and response text alongside the live conversation. |
| Background-noise handling | The model is designed to distinguish pauses, surrounding speech and conversational interruptions more intelligently. |
| Telephony support | Developers can deploy GPT-Live-1 into phone-based workflows such as support, bookings and order handling. |
| Voice-layer price | $0.05 per minute, before backend reasoning and tool costs. |
The bigger change is separating conversation from intelligence
Traditional AI voice agents commonly behave like a chain. Audio is transcribed, the transcript is sent to a language model, it generates a response, and then another system turns that response back into audio.
That architecture works, but every hand-off creates another opportunity for delay or lost conversational context. More importantly, the application has to decide when the person has finished talking before the rest of the chain can reliably start.
GPT-Live-1 moves those interaction decisions into the voice model itself. It can keep listening while generating audio, respond to a short acknowledgement differently from a new request, and allow someone to change direction without waiting for a long answer to finish.
The deeper reasoning is then delegated elsewhere. A business could use a cheaper, faster backend for routine order updates but route a difficult support question to a stronger reasoning model. The voice presented to the customer does not have to change simply because the computational work behind it does.
This effectively turns the voice model into an interface and orchestration layer rather than requiring it to be the smartest model in the stack. For developers building agents at scale, that separation is useful because conversational quality, reasoning quality, and cost can be tuned independently.
$0.05 per minute is not the real cost of a GPT-Live-1 agent
The headline price works out to $3 per hour of conversation for GPT-Live-1 itself. One hundred hours would therefore cost $300 for the voice layer, while 1,000 hours would cost $3,000.
Those figures are only the starting point. Backend reasoning is charged separately, as are tools and any surrounding infrastructure. A voice assistant that mainly confirms appointments could therefore have a very different cost profile from one that continuously invokes a frontier reasoning model, searches databases, and modifies customer records.
Teams should model cost by completed workflow rather than comparing the $0.05 figure with a basic speech API. If all you need is transcription, a dedicated speech-to-text route will usually be a cleaner benchmark. Our OpenAI Whisper API pricing guide shows how different the economics become when the job stops at turning audio into text.
Interruptions create a new implementation problem developers should not miss
Natural interruption handling sounds entirely positive until the voice agent is connected to tools.
Imagine asking an agent to move an appointment, then interrupting halfway through because you remembered another date would be better. The spoken conversation may adapt immediately, but work already delegated to the backend needs its own cancellation or supersession logic.
OpenAI has specifically noted in its developer announcement that interrupting the spoken response does not automatically cancel backend work. Developers therefore need to treat conversation state and task state as separate things.
For any workflow that writes data, spends money, submits an order, or contacts another person, the safer design is to identify the current intent, invalidate obsolete work, and require confirmation for consequential actions. Otherwise, the conversational layer can feel perfectly responsive while an older request is still executing underneath it.
More human backchanneling can actually make voice agents worse
Another less obvious issue is how often the agent should acknowledge the person speaking. GPT-Live can use short responses, pauses and listening cues in ways that make conversation feel less mechanical. Used badly, the same behaviour becomes irritating.
Early reactions to GPT-Live repeatedly focus on agents speaking too eagerly, adding unnecessary verbal acknowledgements or jumping into pauses that a person intended to use for thinking. In other words, producing more human conversational behaviour does not automatically produce a better interface.
Developers should test silence tolerance, response length and acknowledgement frequency with the same care they give latency. A customer-support agent probably needs a different conversational rhythm from a language tutor or hands-free coding assistant.
OpenAI says an early evaluation by language-learning company Speak reduced interruptions during thinking pauses by almost 80% compared with previous turn-based systems. That is encouraging, but it is still a vendor-supplied evaluation. The more meaningful test will be whether developers can maintain that behaviour across different microphones, accents, noisy rooms and long sessions.
GPT-Live-1 does not replace normal AI voice generators
For prerecorded narration, adverts, podcasts, character dialogue and video voiceovers, most teams still need conventional text-to-speech rather than a permanent two-way session.
Those tools are optimised around different constraints, including voice consistency across multiple takes, editing controls, cloning, commercial licensing and export quality. Our comparison of the best AI voice and audio tools covers that category separately.
GPT-Live-1 becomes more interesting when speech is itself part of the control system: phone support, tutors, assistants, accessibility products, interactive characters, meeting agents and applications designed to operate without a keyboard.
What developers should test before moving a production voice agent
- Interruption recovery: change a request while the model is already speaking and confirm that both the conversation and any delegated work end up in the correct state.
- Thinking pauses: test users who pause for several seconds mid-sentence rather than speaking in perfect prompt-shaped turns.
- Background speech: run sessions with televisions, colleagues, traffic and side conversations rather than a quiet development environment.
- Backend routing: measure how often simple requests are being sent to expensive reasoning models unnecessarily.
- Action confirmation: prevent conversational ambiguity from becoming an unintended purchase, booking or account change.
- Long-session cost: calculate voice, reasoning and tools together rather than modelling the $0.05-per-minute figure in isolation.
- Voice behaviour: tune silence, acknowledgement frequency, pacing and answer length for the specific job instead of trying to maximise how human the agent sounds.
DIY AI take: voice is becoming an agent interface, not a separate feature
GPT-Live originally reached ChatGPT users in July 2026. The important development on 10 September is that developers can now put the same interaction model inside their own products.
The release also points towards a different way of building voice software. Rather than choosing one model and expecting it to handle listening, speaking, reasoning and actions equally well, developers can optimise each layer separately. GPT-Live-1 handles the human conversation. Another model can think. Tools can act.
The limitation is that modular systems still need coordination. An agent that sounds natural can still call the wrong tool, continue an obsolete task or become expensive if every spoken question triggers heavyweight reasoning.
So the most interesting question is not whether GPT-Live-1 sounds more human than the previous generation. It is whether developers can use full-duplex speech without losing control of state, permissions and cost behind the conversation. If they can, voice agents become much more practical than the stop-start assistants most people have used so far.
