OpenAI Releases GPT-Live-1.1 in API
The full-duplex speech model can listen while it speaks and is now available under the model ID gpt-live-1.1.
Updated on Nov. 21, 2026 in Artificial Intelligence

OpenAI has made GPT-Live-1.1, a full-duplex speech model for multilingual conversations, available in its API under the model ID gpt-live-1.1. The company also added a standby billing state that charges less for idle voice sessions.
Why it matters
The GPT-Live family replaces the traditional cascade of speech-to-text, a language model, and text-to-speech with a single model, which could change how developers build voice applications worldwide. Lower charges for idle time may also affect the cost of running always-on voice services for organizations.
Standby mode bills idle WebRTC time at $0.02 per minute, 60% less than the $0.05 per minute charged for active time, and live sessions hold a 128,000-token default context window with instructions capped at 16,384 tokens. OpenAI has not published benchmark scores for the model, and it lists four areas where it expects future gains.
The players
OpenAI
An American artificial intelligence research company that develops the GPT family of large language and speech models and offers them to developers through its API.
The details
GPT-Live-1.1 can listen while it speaks and supports multilingual conversations, but it does not accept image, video, or structured outputs. The model runs only on OpenAI's cloud and is accessed through the v1/live/sessions endpoint. Standby pauses inference during silence while the WebRTC connection stays open.
Timeline
September 10, 2026: GPT-Live-1 reached the API.
October 15, 2026: GPT-Live-1 began powering ChatGPT Plus and Pro voice.
November 2, 2026: Outbound SIP calling reached all organizations.
November 21, 2026: GPT-Live-1.1 was made available in the API.
Roadmap
Voice assistants have long been built from separate speech-to-text, language model, and text-to-speech components chained together, which adds latency and turn-taking delays. GPT-Live-1.1 replaces that chain with a single full-duplex model that can listen while it speaks, a shift that moves the category toward more natural real-time conversation.
Developers building voice products can now run idle sessions at a lower rate, which can reduce the cost of keeping a voice connection open between interactions. The model's lack of image or structured output support means teams needing those features will still need other tools.
The takeaway
Developers evaluating voice features should weigh the lower standby rate against the absence of published benchmarks before committing to GPT-Live-1.1 for production use.
Further reading
For more on where voice models are heading, see Artificial Intelligence coverage across the site.
Source note: This article includes information reported by MarkTechPost.



