OpenAI Releases GPT-Live-1 Voice Model in API
The full-duplex voice layer costs $0.05 per minute and processes audio natively while delegating reasoning to separate backend models.
- Improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1.
- Enables mid-turn interruptions, ambient noise filtering, and tone mirroring in a single model.
- Delegates deep reasoning and tool execution to models like GPT-6 Astra, Luna, or third-party APIs.
- Provides native ASR transcripts, response text, keyword biasing, explicit turn detection, and telephony deployment support.
- Allows developers to configure tone, pacing, expressiveness, language, and response length via system prompts.
Developers can deploy interruptible, full-duplex voice agents without orchestrating separate speech-to-text, reasoning, and text-to-speech pipelines.
