OpenAI released GPT-Live-1 and GPT-Live-1 mini today, a pair of full-duplex voice models that can speak and listen at the same time, and it is retiring Advanced Voice Mode in ChatGPT in favor of the new stack. GPT-Live-1 mini becomes the default for all ChatGPT users, while paid tiers get access to the larger GPT-Live-1. More than 150 million people already talk to ChatGPT through Voice and Dictation, making this the largest voice-model swap OpenAI has ever shipped.
The architectural change is the story. The prior voice mode chained three models together — speech-to-text, a language model, then text-to-speech — which introduced latency and a hard turn-taking pattern. GPT-Live-1 collapses that pipeline into a single duplex model that can be interrupted mid-sentence, sit silently while a user thinks, and hand off harder queries to GPT-5.5 for reasoning, search, or agentic work without breaking the conversation.
OpenAI is pitching the model as a long-form conversational partner rather than a command-and-response assistant. During a press briefing, ChatGPT Voice product lead Atty Eleti said he had held 30- to 40-minute conversations with the model on walks. Because the new voice mode is wired into OpenAI's latest text models, it can also surface information visually when the conversation calls for a chart, image, or link — a hybrid response pattern startups like Monogram, which raised $40 million in seed funding from DST and Lux Capital, have been chasing.
Key facts
- 01OpenAI released GPT-Live-1 and GPT-Live-1 mini, full-duplex voice models that speak and listen simultaneously.
- 02GPT-Live-1 mini replaces Advanced Voice Mode in ChatGPT by default; paid tiers get the larger GPT-Live-1.
- 03The new voice stack routes queries to GPT-5.5 for reasoning, search, and agentic tasks mid-conversation.
- 04More than 150 million people already use ChatGPT's Voice and Dictation features.
- 05Product lead Atty Eleti says he has held 30- to 40-minute conversations with the model on walks.
The competitive frame matters here. Apple and Amazon have both retooled their assistants over the past year to hold context and speak more naturally. Sesame, founded by Oculus co-founder Brendan Iribe and Ankit Kumar, is building AI assistants that keep the conversation going while executing tasks in the background. OpenAI is targeting the same problem — hands-free, long-running voice — but with the reach of a product 150 million people already open.
Eleti framed voice as a strategic bet, not just a feature refresh. He tied the new model directly to the agentic work users already do through Codex and ChatGPT, suggesting voice is meant to become the interface layer on top of long-running tasks rather than a chat gimmick.
“Over time, we think this will also unlock the ability to use voice as a kind of primary interface to computing, and to manage increasingly complex long-running agentic work.”— Atty Eleti, ChatGPT Voice product lead at OpenAI
Hardware is the obvious next shoe. Reports this year have suggested OpenAI is building a pair of AI-capable earbuds, and a full-duplex model with 40-minute conversational endurance is the sort of software that only makes sense paired with always-on audio hardware. OpenAI declined to comment on hardware plans during the briefing.
The company was careful to say GPT-Live-1 is not built to be a companion product. Safeguards route teen accounts to age-appropriate responses, and conversations touching self-harm surface external resources. That framing — natural but not intimate — is a deliberate line, and one competitors have crossed with mixed results.
The rollout has rough edges. In a demo of live translation into Hindi, the model spoke with a heavy American accent and produced Hindi that was stilted and bookish. OpenAI said the model is optimized for "most spoken languages" but didn't publish a list. Accent and prosody in non-English languages is a known weak spot for voice models, and OpenAI's demo confirmed it hasn't been solved. Those are early-version quirks and the kind of thing that improves quickly with more training data and fine-tuning passes; the underlying architecture change is the durable news.
For OpenAI, the strategic calculation is that voice is the layer where the ChatGPT install base compounds hardest. Text chat is a crowded market where Anthropic, Google, and open-source models trade blows on benchmarks. Voice, once it works well enough for 40-minute walks, is stickier — it becomes muscle memory, and it opens the door to hardware OpenAI has been quietly building toward. GPT-Live-1 is the software half of that bet, shipping first.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
