OpenAI's ChatGPT release notes say voice is now powered by a new model line, GPT-Live-1 for paid users and a mini version for free users. The notable change is that the models listen and speak simultaneously rather than strictly taking turns, and they run inside the normal chat with streamed text responses, web search and memory available. Separately, paid and enterprise users can now save up to 5,000 characters of custom instructions, up from 1,500. Neither is a flashy launch, but both target the same complaint: assistants that forget your preferences and force stilted turn-taking. The longer instruction budget in particular is the cheapest way to make a general model behave like a specialized one.