← Back to feed

OpenAI shipped GPT-Live on July 8, replacing ChatGPT’s existing voice mode with a full-duplex architecture that can listen and speak at the same time. The rollout covers two models — GPT-Live-1 and GPT-Live-1 mini — going to ChatGPT users globally today, with API access announced as forthcoming.

What Full-Duplex Changes

Voice AI has moved through two distinct architectures since ChatGPT Voice launched in 2024. The original was cascaded: speech-to-text, then LLM, then text-to-speech, three models chained in sequence. The next generation collapsed that into a single turn-based model that processed audio natively, reducing latency and making conversations smoother, but still required the model to wait for the user to finish speaking before responding.

GPT-Live is full-duplex. The model processes user audio and generates output audio simultaneously. It can inject conversational continuers — “mhmm,” “yeah,” a brief pause filler — while the user is still speaking, then continue when the user stops. Short back-and-forth exchanges no longer require waiting for one side to finish before the other can start.

OpenAI says this eliminates the false-turn-detection problem that plagued Advanced Voice Mode: silence detection triggering a response mid-thought. With GPT-Live, the model maintains continuous context of both streams rather than watching for pauses.

Search and Background Reasoning

GPT-Live-1 can hand off complex queries to a frontier model running in the background. At launch, that backend is GPT-5.5; OpenAI says it will update the backend as new models are released. The handoff is transparent to the user in the sense that conversation can continue while the backend processes a web search, a multi-step calculation, or a document analysis. GPT-Live-1 bridges the gap and brings the result back into the voice stream when it is ready.

The architecture separates the real-time conversational model from the reasoning model, keeping latency low for the voice layer while preserving access to frontier-class intelligence for tasks that need it.

Competitive Landscape

ByteDance shipped Seeduplex, its own full-duplex voice system, in May 2026. Google’s Gemini Live 3.1, which leads the Full-Duplex Bench v3 on speed while going silent 22% of the time, set the performance baseline. xAI’s Grok Voice Think Fast 1.0 entered the space in June. GPT-Live-1 is OpenAI’s response — the last major frontier lab to ship a full-duplex voice product, but the first to pair it with integrated search at launch.

OpenAI has not published arena-style benchmarks for GPT-Live-1. The Full-Duplex Bench v3 comparisons against Gemini Live, Seeduplex, and Grok Voice are the most likely comparison point; those results are not yet available for the new model.

API availability is forthcoming. Developers and enterprise customers can sign up for early access. Pricing has not been announced.

Infrastructure Note

GPT-Live sits as a voice layer above OpenAI’s model stack rather than as a standalone model. Its computational cost comes primarily from running the real-time audio processing model continuously, with frontier model calls billed separately when triggered by search or complex reasoning requests. The billing model for API access will determine how competitive GPT-Live-1 is against Gemini Live’s existing API pricing.