OpenAI Launches GPT-Live-1, Full-Duplex Voice API at $0.05/min

OpenAI has introduced a new API, GPT‑Live‑1, designed for simultaneous voice interactions. The company highlights improvements in latency, interactivity, and tool accuracy based on its own tests, while Yelp reports using it for its phone reservations.

Increased Scores on OpenAI's Internal Tests

According to evaluations conducted by OpenAI, GPT‑Live‑1 achieves a full-duplex interactivity score of 80.1%, compared to 45.4% for GPT‑Realtime‑2.1. The measured speaking latency has decreased from 1.4 seconds to 0.8 seconds. OpenAI also reports a tool call accuracy of 87%, up from 60% for the previous model. In a benchmark for banking voice support, the success rate rises to 32% for GPT‑Live‑1, compared to 12.4% previously.

Adoption, Pricing, and Flexibility for Developers

Yelp is utilizing GPT‑Live‑1 for its phone reservations, and according to its CTO Alex Levy, this allows for better call management. Access to the API is priced at $0.05 per minute. Developers have the option to connect GPT‑Live‑1 to various backend models depending on the task, while adjusting the level of reasoning, execution speed, and pricing for each use case. OpenAI indicates that further details will be provided in the API documentation.

Delivered Features: Full-Duplex, Voice, and Transcriptions

GPT‑Live‑1 enables simultaneous voice exchanges through full-duplex mode. The model offers twelve new voices covering various accents, dialects, and languages. It provides ASR transcriptions and response text by default. GPT‑Live‑1 is already integrated into ChatGPT. This voice synthesis model is accessible to developers via an API.