Multilingual pronunciation and transcription improved
OpenAI released the voice model "GPT-Live-1.1" in its developer API on Nov. 19 (Nov. 20 Japan time). The API changelog says the model improves pronunciation, accent consistency and transcription in multilingual voice conversations. The company's developer account, "@OpenAIDevs," announced it on X.
GPT-Live is a voice model that can speak while listening to the user. OpenAI added it to ChatGPT in July and released the first version, "GPT-Live-1," in its API on Sept. 10 (Sept. 11 Japan time). ChatGPT was reported to have switched to 1.1 on Oct. 15, and the same model version is now open to outside developers.
The model's official page describes GPT-Live-1.1 as a voice model for multilingual conversations. Compared with GPT-Live-1, it transcribes speech more accurately and keeps its accent and voice more consistent across languages, according to the page. It takes and produces audio and text, and does not support images or video.
According to the changelog, developers can use it by changing the model name to "gpt-live-1.1" in an existing session configuration. OpenAI advises describing in plain language when the model should hand work to other models or tools on the backend.
On the same day, OpenAI published a developer blog post on writing prompts that match regional ways of speaking. Its list of built-in voices includes the Japanese voices "Harema" and "Shida" and the Hindi voice "Nira." The post does not mention GPT-Live-1.1.
Idle time billed at $0.02 a minute with inference paused
In the same changelog entry, OpenAI made it possible to configure "standby," which pauses inference when no one is talking and resumes when someone speaks. It applies to GPT-Live-1 and GPT-Live-1.1, and only to WebRTC connections.
According to the guide, a session enters standby after a set number of seconds, from 5 to 60, without an input or output transcript. The connection and session ID stay the same during standby, and when the user starts speaking the session resumes with the conversation history restored. If no timeout is set, inference stays active from the first wake until the session ends.
The official model page lists the regular rate at $0.05 a minute and the standby rate at $0.02 a minute. Time spent connected and waiting for the user to speak costs $0.03 a minute less than regular time.
The guide cautions, however, that after standby the first response to a short utterance may take longer. The microphone track must stay active during standby. For long periods without use, the guide also describes closing the session and starting a new one when the user returns.
No performance figures; reliability claim unconfirmed
The GPT-Live-1.1 model page lists performance and speed as "Not specified." As of Nov. 21, the official documentation did not include figures or benchmarks showing how much transcription accuracy or accent consistency improved.
The summary of the announcement that ORIENS recorded during its monitoring cited improved reliability in completing tasks. The changelog, model page and guide contain no statement to that effect, and ORIENS could not confirm it. ORIENS could not check the text of the original post.
Limits on concurrent sessions are counted separately for GPT-Live-1 and GPT-Live-1.1. For GPT-Live-1 the limit ranges from 50 to 500 depending on the usage tier, and the model is not available on the free tier. The 1.1 page gives no specific figures.
Same price as GPT-Live-1, 3 cents a minute below rivals
GPT-Live-1.1 is billed by session duration, at $0.05 a minute for regular time and $0.02 a minute in standby. GPT-Live-1 is also $0.05 a minute regular and $0.02 a minute in standby, so switching to the new version does not raise the price. Calls to other models or tools on the backend are billed separately.
Compared with the rates of other companies' voice-agent models that ORIENS checked as of Nov. 5, it is $0.03 cheaper than the $0.08 a minute of SpaceXAI's (formerly xAI) "Grok Voice 3.0" and ElevenLabs' "Eleven Realtime v1." What the rate covers differs by company, however; ElevenLabs, for example, adds language-model token charges separately.






