語音 AI／開發者平台
GPT‑Live‑1 API Decouples Full-Duplex Voice from Backend Reasoning, Pricing the Voice Layer at $0.05 per Minute
OpenAI is giving developers access to GPT‑Live‑1, which can listen, speak, and handle interruptions simultaneously, while allowing the voice front end to delegate complex reasoning and tool operations to other models. The new architecture reduces fragile handoffs between STT, LLM, and TTS, but backend reasoning, tools, and call infrastructure are billed separately.
OpenAI made the [GPT‑Live‑1 API](https://openai.com/index/introducing-gpt-live-1-in-the-api/) available on September 10. Rather than wrapping an existing three-stage pipeline of speech recognition, a text model, and speech synthesis into a single endpoint, it uses one front-end model to process input and output audio simultaneously. This enables it to interpret pauses, backchannels, background conversations, and users changing course mid-utterance. When a task requires retrieving information, calling tools, or extended reasoning, the front end can delegate the textual context to GPT‑6 Astra, other OpenAI models, or third-party backends while the voice interaction continues.
This separation allows engineering teams to independently choose a low-latency voice layer and reasoning layers with different cost profiles. It also avoids having to reestablish synchronization among ASR, conversation state, and TTS after every interruption. The API natively provides recognition transcripts, response text, keyword biasing, and turn detection, and it supports telephony use cases. OpenAI reports that GPT‑Live‑1 scores 30 percentage points higher than GPT‑Realtime‑2.1 on Full Duplex Bench. When paired with Astra medium, it also ranks first on Tau3’s end-to-end voice-agent benchmark. However, these are vendor-run controlled evaluations involving different backend models and reasoning effort, so the gains cannot be attributed solely to the voice model.
The voice layer costs $0.05 per minute, with backend models and tools billed separately. A [third-party technical overview](https://www.orcarouter.ai/blog/gpt-live-1-api-release) also notes that it uses a separate Live endpoint, meaning existing Realtime applications cannot migrate simply by replacing the model name. Deployers should pay particular attention to task-cancellation semantics after an interruption: stopping spoken output does not mean a backend tool has stopped running. Key issues to watch next include turn detection for Traditional Chinese and mixed-language speech, latency under telephone-network jitter, and the actual per-call cost created by prolonged silence and waits for backend processing.