OpenAI has opened GPT-Live-1.1, the voice model behind ChatGPT Plus and Pro voice since October 15, to developers in its API, according to the OpenAI API changelog. Alongside it comes an opt-in standby mode that bills silent stretches of a call at $0.02 per minute instead of the $0.05 active rate listed on the pricing page. It matters most to teams building voice agents and automated call lines, and Arabic builders should know there is still no ready-made Arabic voice.
What changed
Announced on the changelog on November 19 and the OpenAI developer forum on November 20 under model ID gpt-live-1.1, the model operates in full duplex, listening while speaking. As noted in the model documentation, it can route complex tasks to a chosen backend model. OpenAI claims better accent consistency across languages, sharper transcription, superior handling of letters and numbers, and more reliable instruction following. However, MarkTechPost noted that OpenAI published no benchmark scores, leaving these as unverified claims. The model supports only audio and text—without video, images, or structured outputs—and switching models requires opening a new session with separate concurrency limits.
Tired of juggling ten tabs? ToolSuite bundles the AI workflow tools power users rely on — in one place.
Try ToolSuite NowWhat it costs
Active usage costs $0.05 per minute, billed per second, with backend models and tools billed separately. Under the live conversations guide, the new standby mode works across WebRTC sessions for both GPT-Live-1 and 1.1. Developers configure an idle timeout of 5 to 60 seconds; silence pauses inference while preserving session history until speech resumes. As noted by The Defiant, this opt-in feature costs $0.02 per minute, cutting idle charges by 60%, though first responses after standby can be slower.
| Session State | Price per Minute | Ten-Minute Call |
|---|---|---|
| Active Voice | $0.05 | $0.50 without standby |
| Standby Mode | $0.02 | $0.38 with four idle minutes |
Prices checked on 22 Nov 2026. The example is our own arithmetic: in a ten-minute call with four minutes of silence, standby cuts the voice-session bill from $0.50 to $0.38, before backend model fees. The saving is real for long calls full of waiting, such as a caller hunting for an order number; a short, busy call barely changes.
Who it affects
The release directly affects developers building voice agents and operating call centres with intermittent silence. Teams that today price phone agents on platforms such as Vapi now have a cheaper idle rate to compare against. For consumers nothing changes: ChatGPT Plus and Pro subscribers have been talking to this model since October 15. The news is that outside developers can now build on it directly.
What it means for Arabic speakers
The guide’s list of 21 additional GPT-Live voices covers English, Portuguese, German, French, Hindi, Japanese and Korean; none is Arabic. The workaround is prompt-to-voice, which builds a voice from a text description: it is generally available to all API customers, free to create, and billed at normal session rates when used. A tutorial on the developer blog provides a worked example for Emirati Arabic from Dubai, combining an urban accent description with instructions to use terms like شو, الحين, وين, and مب, and switch between Arabic and English when the caller switches. OpenAI admits that naming the language alone leaves a lot unspecified (an Egyptian or Moroccan caller expects something quite different), and suggests testing an Arabic sentence with an English product name and a price.
Voice cloning from audio remains restricted to sales-approved customers, and Arabic is missing from the 16 languages authorized for recording consent in the custom voices guide. Furthermore, OpenAI published no Arabic quality figures for the model. On regional availability, the supported countries list includes Arab nations like Saudi Arabia, UAE, Egypt, Jordan, Morocco, Algeria, Tunisia, Iraq, Kuwait, Qatar, Bahrain, Oman, Lebanon, Libya, Palestine, Mauritania, Sudan, and Yemen, though Syria was not found on the list. Pricing is in US dollars everywhere, with no regional rates. If you want ready-made tools that already handle dialects, see our guide to AI customer service tools that handle Arabic.
What to do now
Teams preparing to adopt the model should take three practical steps:
- Switch the model ID: Start a new Live session with gpt-live-1.1 at the v1/live/sessions endpoint while managing separate concurrent-session limits.
- Enable standby with an idle timeout: Opt into WebRTC standby with an idle timeout between 5 and 60 seconds to halt inference during conversation silence.
- Test your Arabic before launch: try an Arabic sentence with an English product name and a price, then a caller who switches dialect mid-call. Check the number, currency and product name separately, and time the first reply after standby.

Comments
0No comments yet.
Please log in to comment.
Comments are available for members only. Sign in to participate in the discussion, or create a new account for free.