OpenAI has opened its full-duplex voice model GPT-Live-1.1 to external developers through the Live API, under the model ID gpt-live-1.1, at the v1/live/sessions endpoint with WebRTC or WebSocket connections. The model — which has powered ChatGPT Plus and Pro voice interactions since October 15 — listens while it speaks, accepts audio and text as input, produces audio and text as output, carries a 128,000-token default context window, and supports streaming and function calling. Image and video inputs, structured outputs, fine-tuning and predicted outputs are not supported, and it runs only in OpenAI's cloud.
The more interesting part is the billing. A new standby state prices idle time at $0.02 per minute, down from the $0.05 active rate — a 60 percent cut for sessions that stay connected through silence. Standby pauses inference after a configurable timeout while preserving the WebRTC connection, the session ID and the conversation history, so the agent resumes with full context when the user speaks again. Configuration takes three parameters: wake_policy.type set to speech, idle_timeout_s between 5 and 60 seconds, and an optional vad_threshold between 0 and 1 that defaults to 0.5.
The constraints matter as much as the discount. Muting the microphone does not trigger standby — the session keeps billing at the active rate until the idle timeout elapses without detected speech, and the microphone track must remain active throughout standby. A 15-second startup charge applies when creating a session, and OpenAI warns the first response after a short utterance may carry extra latency. For the target deployments — kiosks, in-car assistants, ambient support widgets that hang on the line for hours — the math shifts visibly: an eight-hour session with two hours of active speech costs about $24 fully billed and roughly $13 with standby enabled.
Quality claims come without receipts. OpenAI says GPT-Live-1.1 has more natural and consistent accents across languages, better transcription, better handling of letters and numbers, and more reliable task completion than GPT-Live-1 — but it has published no benchmark scores, word error rates or listener studies, so all of those gains are vendor-claimed. There are no open weights. European data residency currently lists only GPT-Live-1, not 1.1, and concurrent-session limits for the two versions are separate and undisclosed.
The lineage gives the release context. GPT-Live-1 reached the API on September 10, 2026, and was quickly put to work — Yelp used it for phone reservations, and early developer evaluations from the language-learning startup Speak reported almost 80 percent fewer interruptions than with previous turn-based systems. OpenAI's follow-on Realtime 2.1 release showed what measurable voice improvements look like, with caching cutting p95 latency by at least 25 percent. Backend model and tool usage is billed separately from the voice session, so a voice agent that delegates reasoning to a frontier model pays token charges on top of the per-minute rate.
Pricing silence is a quiet signal about where voice AI is going: the bottleneck for always-on voice agents is not intelligence but the cost of holding a line open all day. At 2 cents a minute for standby, OpenAI is making persistent voice infrastructure — receptionists, kiosks, in-car copilots — economically plausible, and forcing competitors to decide whether silence should be billed like speech.
Comments (0)
Log in to join the discussion
Log InNo comments yet