AI news · 2 min read
GPT-Live can ask Decisions to act while you are still speaking
OpenAI’s October 15 API update connects Live to Decisions. Build a small reversible voice action and account for separate Decisions and Live usage.
Announcement date:
Affiliate disclosure: Some links may earn CreatorStack a commission at no extra cost to you.
At a glance
The October 15 API update lets GPT-Live call Decisions directly or trigger it from transcript changes. The application still carries out and verifies actions. Decisions pricing is separate from the cost of the Live conversation; this is not the ChatGPT Voice model rollout.
A voice-controlled editing tool should be able to prepare a cue without waiting for a whole speech to end. OpenAI’s October 15, 2026 API update adds a native connection between GPT-Live and Decisions for that kind of small, structured response. This is a developer integration; it is separate from the new ChatGPT Voice models. API release entry.
Choosing an action is only part of doing it
Developers define questions and permitted choices. Live can delegate directly to Decisions, avoiding a round trip through the application, or the application can use triggers tied to incoming or outgoing transcript changes. The application still executes the selected action. OpenAI’s guide returns the observed result before Live reports success. Native setup requires the api.decisions.write permission. Integration guide.
An illustrative creator prototype might select a preview cue while someone describes the next shot. Begin with preview selection, which is easy to reverse. Include a correction such as “the second shot—actually, the third” and check whether the visible selection agrees with the final request. Keep publishing and destructive edits behind an application review step. These are proposed design checks, not results from a tested application.
The distinction matters: a fast choice that leaves the wrong clip selected has not saved useful time. Measure the state your interface actually reaches, alongside the wait the person experiences.
Budget for the whole conversation
The Decisions guide lists $0.10 per million input tokens for gpt-6-luna on /v1/decisions, with no output, cache-read or cache-write charges. Regional processing premiums and long-context multipliers apply. That endpoint rate does not describe the whole Live session or every request made with the same model. Pricing and availability.
Our Live API explainer covers the surrounding voice architecture. Use the AI output evaluation guide to define an acceptance check before giving your prototype more consequential controls.
Common questions
Questions people ask.
- Does a decision result prove that the requested action happened?
- No. Your application must perform the action and return its observed result. A model selecting an action is not evidence that the application completed it.
Sources & freshness
This article was refreshed on 16 Oct 2026 using the sources below. A source check verifies published information, not output quality or checkout availability in every region.
- developers.openai.com/api/docs/changelog
- developers.openai.com/api/docs/guides/decisions-voice
- developers.openai.com/api/docs/guides/decisions
What changed: Original coverage of the October 15 announcement with product, access and practical limits checked October 16.
Found a correction? Let us know.