OpenAI Launches GPT-Live-1 on API with Full-Duplex Voice Conversation Support

OpenAI launched GPT-Live-1 on its API on September 10, 2026, opening up the natural full-duplex voice conversation features previously available in ChatGPT to developers. This model processes incoming and outgoing audio simultaneously with a single model, eliminating the latency and instability of existing STT-LLM-TTS pipelines.

Performance and Architecture Improvements

GPT-Live-1 showed a 30%p performance improvement over GPT-Realtime-2.1 on the Full Duplex Bench, with significant improvements in turn-taking latency and interaction behavior. Interruptions decreased by approximately 80% compared to previous systems. Additionally, it ranked first on the Tau3 benchmark when paired with GPT-6 Astra (medium reasoning effort), demonstrating end-to-end task success rates (Pass@1) in aviation, retail, and telecommunications domains.

In terms of architecture, the codebase was simplified by 80% compared to cascaded builds, with 23K lines of code removed, significantly reducing development complexity. Developers can ensure flexibility by selecting backend text models such as Luna (for high-frequency tasks) or Astra (for complex reasoning) based on task complexity.

Integration and Use Cases

GPT-Live-1 allows control over tone, speed, and style via system prompts, with improved handling of background noise and silence. Major companies such as Yelp, Speak, Fin, and Cognition have applied it to their services, reporting benefits such as improved reservation processing rates, natural conversation flow, and enhanced collaboration experiences with AI engineers. Integration with the Codex SDK also supports a structure that passes conversation context to Codex threads and returns responses.

Pricing and Availability

The API is immediately available, priced at $0.05 per minute for the front-end voice layer. Backend models and agent harnesses must be paired separately according to product requirements. Voice options including a wider variety of accents and dialects than before are provided, and language availability is expected to continue expanding over the coming months.