GPT-Live can now call Decisions mid-sentence, before you reply

OpenAI integrated its Decisions API into GPT-Live, its realtime voice API, on 15 October. The change lets a voice agent act while someone is still talking, rather than waiting for them to finish and for a full response to come back.

There are two ways to wire it up. With native delegation, GPT-Live calls Decisions directly: you define questions up front in the session, feed it context through decision.item.create, and get answers back as decision.completed events, which your app turns into actions reported via session.commentary.append. OpenAI’s own example is a car voice assistant: “make it warmer” resolves to a cabin temperature action without a round trip through your server. The alternative is event-based triggers configured on session.triggers, which fire off session.input_transcript.delta while the caller is speaking or session.output_transcript.delta while Live is responding, useful for things like syncing an avatar’s expression to the tone of what’s being said.

Why it matters: voice agents built on a request/response loop always have a gap between “user finishes talking” and “agent decides what to do.” Running Decisions inside the Live session itself, rather than as a separate call from your backend, removes that round trip for the classification and routing calls that voice interfaces lean on most.

The caveat: this builds on the Decisions API OpenAI shipped on 6 October, which is still in public beta and restricted to gpt-6-luna.