Thesis: Voice customization is becoming an application-level asset, while product boundaries still determine where that asset can run.
Observed facts: The guide says prompt-to-voice is generally available to API customers and that a prompt-generated voice can be created and saved without a reference recording. Observed facts: These generated voices work with GPT-Live, but not with the Text to Speech, Realtime, or Chat Completions APIs; the guide also specifies a limit of 100 such voices per organization. Mechanism: A description is turned into a reusable saved voice, decoupling voice selection from a one-off speech-generation request, while API compatibility remains the deployment constraint.
Why now: OpenAI has also added nine preset voices made with native-speaking actors across French, German, Hindi, Japanese, and Korean, making voice choice more relevant to teams shipping multilingual interfaces. First implication: Builders should treat voice selection as a GPT-Live-specific design and integration decision, not assume a voice can move across all audio APIs. The guide further requires adult voices, transparency that speech is AI-generated, and preservation of provenance signals. Evidence boundary: The public guide documents supported workflows and policy requirements; it does not establish usage, quality, or performance outcomes. First watch condition: Watch for documentation that extends prompt-generated voices to additional APIs or changes the 100-voice organizational limit.
This starts a distinct analytical record because no prior New Runtime record was supplied for this specific combination of prompt-generated voices, API restrictions, and new international presets.