Vercel's AI Gateway integration lets a voice application use OpenAI's GPT-Live 1 while choosing a separate text model for delegated work. GPT-Live 1 can listen while speaking, allowing a user to interrupt or add information. When deeper processing is needed, the application can call another model and return its result while the voice session continues, according to the integration announcement.
This separates the conversational interface from the model handling background work. The application remains responsible for the delegated task's permissions, confirmations and cancellation. Background model calls are billed separately, and the ongoing voice session continues to accrue usage.
Key takeaways · AI-distilled
Delegated model calls are billed separately through AI Gateway from the underlying voice session, and it's the calling application — not GPT-Live 1 itself — that controls permissions, confirmation, and cancellation for whatever the delegated model does.
The handoff runs on a delegationId: the app gets a delegation-created event from the voice model, generates a response with any text model on AI Gateway, then returns the result tagged with that same delegationId via a context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →-append event so the voice model can speak it.
Delegated results are routed onto a dedicated 'commentary' channel (set via providerOptions.openai.channel) rather than being mixed directly into the main conversation transcript.
Integrating it today requires a specific low-level stack — AI SDK 7, @ai-sdk/openai 4.0.67+, and a raw WebSocket client — with connection setup, audio streamingSending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.Full definition →, and transcript assembly left to the integrator rather than a simpler high-level API.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
