WHAT SHIPPED
● GPT-Live-1, released through the OpenAI API.
● Full-duplex: listens and speaks simultaneously rather than taking turns.
● Handles interruptions natively rather than treating them as errors.
● Delegates deeper reasoning and actions to backend models and tools.
Why the handoffs were the problem
Most voice systems chain three components: speech recognition turns audio into text, a language model produces a response, and synthesis turns that back into audio. Each boundary adds latency, and each one is a place where things break.
Interruption is the clearest example. In a turn-based system, someone speaking over the model is an exception to handle. In conversation between people it is completely ordinary, and handling it badly is what makes voice assistants feel like machines.
THE DELEGATION PART IS THE UNDERRATED BIT
GPT-Live-1 hands deeper reasoning to backend models and tools rather than doing everything itself.
That is the right shape. Conversation needs to be fast; reasoning needs to be good. Making one model do both means compromising on one of them, and previous voice products mostly compromised on the reasoning.
What it competes with
| Product |
Approach |
| GPT-Live-1 | Full-duplex, delegates reasoning to backend |
| Grok Voice | Low latency, 300-500ms, strong on naturalness |
| Gemini Flash Live | Speech-to-speech, Google ecosystem |
| ElevenLabs | Synthesis quality, subscription rather than per-minute |
Pricing was not published in the announcements available at the time of writing. Voice is typically billed per minute rather than per token, so compare on that basis rather than against text model rates.
What you would build with it
- Support lines where callers interrupt. That is most support lines, and turn-based systems handle it badly.
- Hands-free operation — field work, driving, anything where a screen is not available.
- Accessibility tooling, where the latency of turn-taking is the barrier rather than the transcription accuracy.
- Anything that currently uses three separate vendors for recognition, reasoning and synthesis. That stack is the thing being replaced.
What to check before building on it
- Per-minute pricing, which was not published at launch. Voice costs scale differently from text.
- What the backend delegation actually calls, since that is where the reasoning cost sits.
- Latency under real network conditions, not demo conditions. Full-duplex is unforgiving of jitter.
- Whether it handles your languages at the same quality as English.
Sources
FAQ
What is GPT-Live-1?
An OpenAI voice model available through the API, supporting full-duplex operation — listening and speaking simultaneously rather than taking turns.
What does full-duplex mean in practice?
The model can process incoming speech while producing output, so interruptions are handled natively rather than as exceptions. That is how human conversation works and it is what turn-based systems get wrong.
Does it do the reasoning itself?
It delegates deeper reasoning and actions to backend models and tools. Conversation needs to be fast and reasoning needs to be good, and separating them avoids compromising on one.
How much does it cost?
Per-minute pricing was not published in the available announcements. Voice is generally billed per minute rather than per token, so check the pricing page directly.
How does it compare to Grok Voice?
Grok Voice leads on naturalness at 300 to 500ms latency. GPT-Live-1 is an API product for building voice systems rather than a consumer assistant, so they are not directly substitutable.