OpenAI GPT-Live-1 and GPT-Live-1 mini: The Full-Duplex Voice Models That Listen and Speak at the Same Time

• GPT-Live-1 — rolls out to paid ChatGPT tiers (Go, Plus, Pro)
• GPT-Live-1 mini — rolls out to Free-tier users

Both replace the previous ChatGPT Advanced Voice Mode. They are built on a true full-duplex architecture that continuously processes incoming audio while generating spoken responses, rather than the old half-duplex approach where the model had to finish talking before it could hear the user ^HAT.

OpenAI said inference costs have come down enough to make the economics of voice work, which is why the model is also available to free users ^A. At launch, GPT-Live uses GPT-5.5 in the background for reasoning ^OH.

How full-duplex voice improves on turn-based assistants

Full-duplex capability changes the feel of voice AI in three specific ways that OpenAI demonstrated during its launch briefing.

1. Natural interruptions and turn-taking

The model can be interrupted mid-sentence and will stop and listen, rather than talking over the user. It also detects when a user pauses mid-thought and waits instead of cutting in ^AT. OpenAI product lead Atty Eleti said during the briefing: "This is a full duplex model" ^T.

The model can also show it is paying attention with phrases like "mhmm" or "yeah" while the user is still speaking, and it will stay quiet when you need a moment to think ^OB.

2. Routing complex queries to frontier text models in parallel

When GPT-Live encounters a hard question, it delegates reasoning to OpenAI's latest text models (such as GPT-5.5) in parallel, while GPT-Live itself stays in conversation with the user ^AT. Research lead Kundan Kumar explained: "When GPT-Live has to think hard for a question, it can delegate its reasoning and complex task to GPT-5.5 in parallel, while GPT-Live still remains in conversation with the user" ^LT. This means users do not wait for the voice model to reason through heavy tasks — the transition from research to spoken findings is near-instant ^T.

3. AI-generated visuals during conversation

GPT-Live can supplement discussions with on-screen visuals — images, diagrams, or other AI-generated content — creating a multimodal experience that goes beyond pure voice chat ^HAT. The Verge reported that the model will supplement conversations about weather, stocks, and sports with AI-generated images ^T.

Collectively, these features make conversations feel closer to natural human dialogue. Business Insider described it as an assistant that "wants you to talk over it" ^B. The Verge's headline called it "better at shutting up" ^T.

Who led development?

The two key people named in press briefings and media reports are:

• Kundan Kumar — Research lead for GPT-Live, described by The Verge as "OpenAI researcher lead" for the project ^LT. He joined OpenAI from Character.AI in mid-2025 and served as the core lead for the audio AI project ^IIC.
• Atty Eleti — Product lead for ChatGPT Voice, who presented the model during the launch livestream ^AT.

Earlier reports from January 2026 also named infrastructure lead Ben Newhouse and product manager Jackie Shannon as involved in the audio AI restructuring ^IX.

Rollout timeline

• July 8, 2026 — Global launch began rolling out to all ChatGPT users on iOS, Android, and web, with GPT-Live-1 for paid subscribers and GPT-Live-1 mini for free users ^RHT.
• Also on July 7–8, OpenAI released new API Realtime models (gpt-realtime-2.1 and gpt-realtime-2.1-mini) with 25% lower p95 latency through improved caching ^C.

One limitation noted at launch: GPT-Live does not currently support voice with video capabilities ^T.

Other major AI releases the same week

OpenAI launched GPT-Live in the middle of an exceptionally busy week across the AI industry.

OpenAI's own broader ecosystem:

• GPT-5.6 (Sol, Terra, and Luna variants) was expected for public release on July 9, 2026, widening access after a limited earlier rollout ^HE.

Competitors and adjacent news:

• Anthropic — Pushed a significant update to Claude Sonnet 5 in early July, with improved instruction-following and code generation accuracy for complex agentic tasks ^S. The Trump administration also lifted export controls on Anthropic's Fable 5 and Mythos 5 models that same week ^SLF.
• Google DeepMind — Shipped NanoBanana 2 Lite, an image-generation model that produces images in under 4 seconds at $0.034 per 1,000 images ^T.
• Meta — Reported to be preparing to offer neocloud services, and separately under pressure from the U.S. government on AI safety reviews ^NL.
• SpaceXAI (xAI) — Reportedly preparing to release new AI models in partnership with Anysphere (Cursor) as part of a broader competitive push ^E.
• China — A 9-day countdown began to China's new AI companion law, which would force shutdowns affecting 345 million Doubao users ^B.
• White House — Expected to announce voluntary AI model standards ^B.