OpenAI sets December 11 cutoff for two ChatGPT voice modes OpenAI sets December 11 cutoff for two ChatGPT voice modes

A selector changes before every feature moves

The retirement notice separates the model choice from the content people have already created. Existing conversations stay in history and can continue in text or voice. During the transition, users can still select the older modes. Once the deadline arrives, Settings > Voice will stop offering them and previous selections will default to Live. This makes the cutoff a change in how the next conversation runs, rather than a deletion of earlier conversations. The explicit exception for visual input also matters: removing a selectable voice mode does not mean every feature using that model ends simultaneously.

Conversation timing is part of the replacement

GPT-Live arrived on July 8 with a different interaction design. OpenAI describes a full-duplex system that listens while producing speech, deciding repeatedly whether to speak, pause or use a tool. Standard Voice instead passed a turn through transcription, a text model and speech generation. Advanced Voice handled audio directly within a model, while retaining separate speaking turns. The difference affects pauses and interruptions, not only the sound of a voice. GPT-Live also separates conversation from heavier work: another model can search or reason in the background while the spoken exchange continues. That division lets the conversational layer remain responsive during a longer task.

The adoption argument has limits

OpenAI's November 10 release notes say voice usage has nearly doubled since GPT-Live launched and that virtually all voice users now use it. The company presents consolidation as a way to focus development on one experience. It identifies response depth, continuity and personalization as areas it is improving. Those are different measures from the amount of use. More usage describes adoption; it does not, by itself, establish that every established workflow transfers intact. For a person who dictates a long thought or pauses frequently, the timing of a reply can be as important as its factual content.

API builders have a separate calendar

Developers embedding audio in their own software face a different set of dates. OpenAI's July 20 API notice schedules legacy audio and realtime families for removal on January 20, 2027. Its table maps older realtime models to gpt-realtime-2.1 and older audio models to gpt-audio-1.5, with separate replacements for smaller variants. That is an endpoint migration inside an application, not a choice in ChatGPT's settings. A team using both products therefore has two distinct inventories: the ChatGPT experiences its staff rely on and the model identifiers its code actually calls.