October 19, 2026 · Ahmed · More by Ahmed
Native Decisions Delegation Arrives in GPT-Live
OpenAI's GPT-Live can now call the Decisions API directly mid-call. Here's what native delegation handles, and what your application still has to do.
On October 15, 2026, OpenAI gave GPT-Live a direct line to its Decisions API. Instead of a voice session notifying your backend and waiting for it to call Decisions, GPT-Live can now call Decisions itself while the caller is still mid-sentence. For teams building voice-based commerce support or ordering agents, that's a meaningful shift in where the decision actually happens, but it's worth being precise about what moved and what didn't.
Three Question Types, One Typed Answer
Decisions, in public beta since October 6, 2026 with gpt-6-luna as the only supported model, evaluates text or images and returns typed answers about 10x faster than the Responses API. It supports three question types. A predicate question returns a probability from 0 to 1 that a condition is true, useful for something like whether a caller is asking to cancel. A choice question returns one value from a fixed set you supply, such as which department should handle a request. A score question rates an input against ordered levels and returns the probability-weighted average of their indices, so the result can land between two levels rather than snapping to one.
What Native Delegation Changes
Before this release, a voice agent that needed a Decisions-style answer had to use custom delegation: GPT-Live notifies the application, which then calls Decisions and reports back. Native delegation removes that round trip, GPT-Live calls Decisions directly. There are two ways to trigger it. session.delegation lets GPT-Live decide for itself, based on its instructions, when a request matches something Decisions can answer. session.triggers instead fires a decision automatically as input or output transcript deltas arrive, so an app can react to what the caller or the agent is saying in real time. Both are set when the session is created, and the delegation type stays fixed for that session; switching modes means starting a new one.
Decisions Picks the Answer; Your App Still Has to Act
None of this makes the voice layer autonomous. Decisions only selects an answer, it doesn't touch any real system. The documented flow still routes through the application: read the chosen answer off the decision.completed event, execute the actual change, confirm what state resulted, send that confirmed state back with decision.item.create, and only then narrate it to the caller via session.commentary.append. Mid-call, the questions themselves can be changed with session.update, but the app has to wait for the matching session.updated event before adding a new decision item for that step. None of it works without a working Live connection carrying the api.decisions.write permission. For a voice commerce agent, that's the real design takeaway: native delegation shortens the path to a fast typed answer, but execution, confirmation, and reporting back are still entirely the application's job.