What changed
OpenAI began rolling out GPT-Live-1 for paid ChatGPT users and GPT-Live-1 mini for free users on July 8. The models use a full-duplex design: they can process incoming speech while producing speech, rather than treating a conversation as a sequence of neatly separated turns.
That should reduce a familiar failure mode in voice interfaces. A pause no longer has to mean the user has finished, and an interruption can become part of the interaction instead of an error that resets it.
Why the architecture matters
The more consequential design choice is delegation. GPT-Live handles the immediate conversation, while work that needs search or deeper reasoning can be handed to another model in the background. At launch, OpenAI says that role is filled by GPT-5.5.
This separates two jobs with different latency requirements. The interaction layer must react in fractions of a second. The reasoning layer may need far longer. If the boundary works well, voice products can stay responsive without forcing every difficult task through a low-latency model.
What is not proven yet
The published preference and safety results are OpenAI's own evaluations, not independent production benchmarks. GPT-Live also launched without video or screen sharing, and the planned API had not shipped when this article was published.
The next useful receipts will be developer access, measured interruption and latency performance outside demos, and evidence that long-running delegated tasks return the right result to the right point in a live conversation.
Watch the API release and whether developers can inspect, control, and recover delegated background work without breaking the conversation state.
Sources
- 1OpenAI — Introducing GPT-LivePrimary announcement
- 2OpenAI Deployment Safety Hub — GPT-Live system cardPrimary safety documentation
- 3OpenAI Help Center — ChatGPT release notesPrimary availability record
Published after human review of the linked sources. Analysis is informational and is not financial advice.