Definition
GPT-Live is openai’s full-duplex voice model family replacing ChatGPT Advanced Voice Mode (July 8, 2026). Models listen and speak simultaneously, enabling natural interruptions, back-and-forth conversation, and live translation.
Architecture
- Full-duplex: continuous input processing while generating output; decisions many times per second (speak, listen, pause, interrupt, invoke tool)
- Background delegation: complex tasks (web search, reasoning) delegated to gpt-56 while maintaining conversational flow
- Variants: GPT-Live-1 (Go/Plus/Pro default), GPT-Live-1 mini (Free default)
Desktop / Codex Activation (July 23, 2026)
- First native voice activation for codex and chatgpt-work on ChatGPT desktop (macOS/Windows)
- Concurrent agent orchestration by voice; appshots screen context on macOS
- Access: Plus, Pro, Business, Enterprise, Education; voice tasks consume Codex/Work quotas
- Distinct from July 8 model debut — this is developer/desktop integration (voice-agentic-coding)
Launch Limitations
- No video or screen sharing at July 8 launch (desktop later adds Appshots)
- Limited language support; non-native accents possible
- API not available day-one; developer sign-up for notification
Scale
150M+ weekly ChatGPT voice users at model launch; Codex/Work claim 10M+ WAU (company claim) at desktop voice activation.