OpenAI opened prompt-to-voice to every API customer on October 16 and added nine named GPT-Live voices the same day. ChatGPT Voice is rolling out GPT-Live-1.1 and GPT-Live-1.1 Quip, a smaller sibling, with new language voices. AWS put Reflection Beam on SageMaker JumpStart. GPT-Live can now call Decisions while the user is still talking.
In the news
Prompt-to-voice
OpenAI's October 16 API release notes make prompt-to-voice generally available to all API customers. You write a text description, create a custom voice through the API or the voice Playground, and reuse it in GPT-Live sessions. Creating that prompt-generated voice is free. Speech generation is billed as a regular GPT-Live session.1
The same notes add nine named GPT-Live voices: Nuri, Juni, and Flitz in German; Brise and Sillage in French; Nira in Hindi; Harema and Shida in Japanese; and Noeul in Korean. OpenAI calls the new voices natural. That quality claim is OpenAI-reported.
Live-1.1
OpenAI's October 15 ChatGPT notes roll out GPT-Live-1.1 and GPT-Live-1.1 Quip, a smaller sibling, in ChatGPT Voice. The company says accents are more natural across languages, and Live uses a wider set of short phrases while it searches. New Live voices cover French, German, Hindi, Japanese, Korean, and Mexican Spanish, plus Bubbie in English. Which voices you hear depends on region and account.2
This is the consumer Voice increment, not the API prompt-to-voice ship. Accent quality and how Quip compares with Live-1.0 are OpenAI-reported.
JumpStart Beam
AWS's October 15 machine learning blog puts Reflection Beam on Amazon SageMaker JumpStart. Beam is an open-weight sparse mixture of experts at 501 billion parameters with 23 billion active, licensed Apache 2.0, with a reported one million token context. JumpStart lists it in us-east-1 and us-east-2. HyperPod Inference can serve it with the JumpStart container, SGLang, or vLLM. Checkpoints ship as BF16, FP8, and NVFP4.3
This is a cloud landing of an open coding model, not a new model announcement. JumpStart is the managed-endpoint path; HyperPod is the Kubernetes-native path. AWS points FP8 at ml.p5.48xlarge, ml.p5e.48xlarge, and ml.p5en.48xlarge, and NVFP4 at ml.p6-b200.48xlarge. Parameter counts, context length, and instance guidance are Reflection or AWS reported. You pay for the endpoint.
Decisions in Live
OpenAI's October 15 API notes add native Decisions integration in GPT-Live. An app can act while the user is speaking or react to what Live says. Live can call Decisions through native delegation, or trigger them as input or output transcript deltas arrive.4
Native delegation is the direct call from Live. Transcript deltas fire as Live hears or speaks. This is the voice-agent action path, not the ChatGPT Voice model increment. Latency of those mid-utterance calls is OpenAI-documented.
