Changes confirmed high confidence

GPT-Live Launches as ChatGPT's Default Voice, Replacing Advanced Voice Mode for 150M+ Weekly Users

OpenAI's full-duplex voice models make interaction decisions many times per second — and ship with an unusual candor note: small disclosed safety regressions versus the mode they replace.

OpenAI launched GPT-Live on July 8, 2026, making full-duplex voice the default ChatGPT voice experience for its more than 150 million weekly voice users and replacing Advanced Voice Mode, per explainx.ai's launch coverage and SQ Magazine's report quoting OpenAI's announcement.

Context

Voice has been ChatGPT's fastest-growing interface. GPT-Live replaces the pipeline-based Advanced Voice Mode with native full-duplex models that make interaction decisions — when to listen, interrupt, respond — many times per second.

What changed

Why it matters

Voice is becoming an agent control surface, not just an input method. Notably, OpenAI paired the launch with a system card disclosing small, non-significant safety regressions versus Advanced Voice Mode — a candor milestone for a default-feature replacement at this scale. Safety work includes audio-native evaluations for self-harm, emotional reliance and psychosis, with predefined voices only.

Details

BuildFastWithAI's review of the official announcement catalogs five disclosed limitations: no video, language-coverage gaps, no API, the GPT-5.5 backend lock, and ongoing emotional-reliance monitoring.

Rollout detail

The tier split mirrors OpenAI's broader access ladder: paid plans get the full GPT-Live-1, free users get the mini variant. Product lead Atty Eleti featured in launch communications, per SQ Magazine's report. The GPT-5.5 delegation backend means complex queries are effectively routed to a text-class model mid-conversation — a hybrid architecture that preserves voice fluency while borrowing reasoning depth, but also ties voice quality to a previous-generation backend until OpenAI migrates it.

Limitations and caveats

Quality metrics (GPQA, preference rate) are vendor-reported. The safety regressions, while disclosed as non-significant, are described only in OpenAI's own system card; no independent evaluation of GPT-Live's voice-specific safety behavior was available at publication.

Sources

*Update note: This post was last reviewed on 2026-07-22. Watch for API availability and Business/Enterprise tier access.*

Sources

Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.