ChatGPT✓ ReviewedTry itChecked 3 September 2026GPT-Live listens while it talks. Fast answers still deserve slow checks.
OpenAI's GPT-Live voice models hold full-duplex conversation, listening and speaking at once, and quietly hand harder questions to a stronger model in the background across ChatGPT web, iOS and Android.
Bit’s takeaway
What changed
OpenAI introduced the GPT-Live family (GPT-Live-1 and GPT-Live-1-mini), full-duplex voice models that listen and speak simultaneously instead of taking turns, and rolled out across ChatGPT surfaces. They delegate harder tasks to a background reasoning model that OpenAI keeps updating as new flagships ship: GPT-5.5 at launch, though by September 2026 the desktop app's Codex and Work voice path had already moved to GPT-5.6 Terra. Supported generated audio also carries SynthID watermarking under OpenAI's provenance system.
Why it matters
Conversation that flows like a person removes the pause that used to remind you a model was answering. The interface got more convincing, not more correct, and a confident spoken answer is harder to fact-check mid-flow than a paragraph you can reread.
Who should care
- People using voice as their main way of working with AI
- Teams building spoken assistants on OpenAI's stack
What to do
Use voice for exploration, drafting and hands-busy moments, and make one habit non-negotiable: anything factual that leaves your desk gets checked in text against a source, because you cannot reread a conversation that never paused.
The human take
Tools change fast. Your judgment matters more.
The person keeps the habit the interface no longer prompts: pausing to verify before a fluent answer becomes a decision.
Affected guidance
Put this to work
Verified facts
- OpenAI introduced the GPT-Live voice model family (GPT-Live-1 and GPT-Live-1-mini), full-duplex models that listen and speak simultaneously and delegate harder tasks to a background model that OpenAI continuously updates as new flagships ship. General ChatGPT voice surfaces still ran on GPT-5.5 as of OpenAI's original documentation; by September 2026 the desktop app's Codex and Work voice path had already moved to GPT-5.6 Terra.Checked
Sources and method