OpenAI has released GPT-Live-1.1 (model ID gpt-live-1.1) via its API, a full-duplex multilingual voice model that processes audio input and output in a single stream, replacing the traditional cascade of speech-to-text, LLM, and text-to-speech. The model maintains a 128,000-token default context window, supports streaming and function calling, and delegates complex reasoning and tool use to backend text models such as GPT-6 Astra or Luna, or third-party backends via client or responses delegation modes. Instructions support up to 16,384 tokens. OpenAI claims improvements over GPT-Live-1 in accent consistency, transcription accuracy, letter and number handling, and task completion reliability, though no benchmark scores were published for version 1.1. A new standby mode reduces idle WebRTC billing from $0.05 to $0.02 per minute (a 60% reduction) by pausing inference during silence while keeping the session and history intact; it activates after a configurable idle timeout of 5 to 60 seconds using a speech wake policy with optional VAD threshold. Billing is per second for voice duration, while backend model and tool usage are billed separately. The model runs exclusively on OpenAI's cloud via v1/live/sessions over WebRTC or WebSocket, with no open weights or quantized variants. Outbound SIP calling is supported. Image, video, and structured outputs remain unsupported.
Read original
marktechpost