Vibeleaderboard
← All Intel
Intel / article

GPT-Live 1.1 now available on AI Gateway

Source
vercel.com
Author
Kevin Dawkins
Date

Context

Vercel's AI Gateway, a service that lets developers call models from many providers through one API and one bill, now offers OpenAI's GPT-Live 1.1 (changelog). For an application that already uses the first GPT-Live model through the gateway, the upgrade is a change of model ID to `openai/gpt-live-1.1`, because the new version uses the same Live endpoint and the same client delegation. Vercel says 1.1 transcribes speech more accurately than the previous model and keeps a more consistent accent and voice across languages, without giving figures.

GPT-Live is a full-duplex voice model: it listens while it speaks, so a caller can interrupt, pause or add detail in the middle of a reply. When the first model arrived on the gateway in September, Vercel contrasted this with voice models that rely on turn detection to decide when to respond. Client delegation is how the voice model gets help with work it cannot do by talking. Session instructions tell it when to ask; it signals a request with a `delegation-created` event, the developer's application answers with any text model on the gateway, and the result goes back on a commentary channel for GPT-Live to say aloud while the conversation continues. Vercel's GPT-Live guide, written for the first model, adds that the delegation event carries an ID and metadata, not a description of the task, so the application has to work out what was asked from the conversation transcript and its own state.

The design leaves control and part of the cost with the developer. The application decides what delegated work may do, including permissions, confirmations and cancellation. Delegated model requests are billed separately through the gateway, and the voice session keeps accruing usage while they run; the guide says voice sessions are billed by duration, including silence and time spent waiting for backend work. The changelog does not list a price for 1.1, saying only that the gateway passes on provider pricing without markup.

Key takeaways · AI-distilled
  • GPT-Live 1.1 is a full-duplex voice model: it listens while it speaks, so a caller can pause, interrupt or add detail mid-reply. Vercel says it transcribes speech more accurately than the previous GPT-Live model and keeps a more consistent accent and voice across languages.
  • Upgrading an existing integration only requires the new model ID, because GPT-Live 1.1 uses the same Live endpoint and client delegation as the earlier GPT-Live model.
  • With client delegation on, GPT-Live sends a delegation-created event when it needs help. The app answers with any text model on AI Gateway and returns the result on the commentary channel for GPT-Live to speak, keeping control of permissions, confirmations and cancellation.
  • Delegated model requests are billed separately through AI Gateway, and voice-session usage keeps accruing while those requests run.
More from Kevin Dawkins
Recommended reads
Comments

Checking sign-in…

Loading comments…