OpenAI added a direct connection between GPT-Live and its Decisions service on October 15, 2026. Teams building voice-controlled apps and conversational characters can now configure action choices or reactions to spoken dialogue within a Live session. OpenAI’s changelog records the update.
Our recommendation for GPT-Live voice assistants is to separate three moments: react, prepare or commit. A character’s nod can change with the conversation; a booking needs a settled choice and any required confirmation.
What the new connection changes
Live can ask Decisions to choose an action when it recognizes a supported request, or trigger decisions as the user’s or assistant’s transcript grows. It calls the service directly instead of asking the app to forward the request. OpenAI’s examples cover vehicle settings and character animation; they are illustrations, not evidence of deployment outcomes. See the integration examples.
Decisions remains in public beta as of October 16, with GPT-6 Luna as its only supported model. It returns fixed choices, probability estimates or rubric scores; an explanation or a multi-step task still needs a suitable backend. OpenAI’s Decisions documentation describes that scope. Our earlier Decisions guide covers choosing between this interface and generated answers.
Choose timing by what the action changes
The following are design recommendations based on the documented workflows, not automatic safeguards supplied by the integration.
Interaction | When to proceed | Control to retain |
|---|---|---|
React as dialogue develops. | Keep the animation reversible, with a neutral fallback. | |
Prepare once enough information is available. | Discard outdated results; a lookup must not silently become a booking. | |
Act on a clear, authorized request. | Apply required confirmation and verify the resulting setting. | |
Confirm the chosen slot before committing. | Keep searching for options separate from creating the booking. |
Preparation can begin without committing the user to an outcome. OpenAI’s delegation guide already describes starting lookups from partial transcripts, so continuous reaction itself is not new. The October 15 change makes direct Decisions calls configurable inside Live.
A selected answer is still only an input to the application. The app checks permissions, obtains required confirmations and runs its functions. OpenAI’s Live overview keeps those responsibilities with the application; a confident classification does not authorize an action.
Handle corrections before reporting success
Transcript-triggered results can arrive out of order. Each identifies the transcript update behind it; the app should apply the newest relevant result for incoming and outgoing speech separately. The integration guide documents this ordering rule.
Imagine a caller asks for Friday, then changes it to Thursday. A Friday lookup arriving later must not replace the corrected request. OpenAI’s task-management guidance recommends tracking request revisions and individual actions, and checking whether an action already happened before retrying it. A delayed response must not create a second booking.
Stopping the assistant’s voice is also different from canceling its task. The backend must process a cancellation and return its outcome; the assistant should not guess what happened while waiting. OpenAI’s prompting guidance makes that distinction.
Before launch, include mid-sentence corrections, delayed results and duplicate-action checks in an evaluation. Inspect the final booking or setting and the spoken confirmation separately. OpenAI’s voice-evaluation guide follows the request through to the application’s final state: a fluent reply alone is not proof of completion.
Budget for conversation and decisions separately
At prices checked on October 16, GPT-Live costs US$0.05 per minute, billed per second. Decisions costs US$0.10 per million input tokens, with no output-token charge. Regional processing premiums and long-context input multipliers can apply to Decisions.
Illustrative calculation, not observed usage: assume one completed ten-minute billable voice session and 200,000 total billed Decisions input tokens across its evaluations, at base rates.
Voice: 10 × $0.05 = $0.50.
Decisions: 200,000 ÷ 1,000,000 × $0.10 = $0.02.
Subtotal for these two components: $0.52.
That excludes other model calls, tools, telephony, hosting, taxes and any extra retry usage; it assumes no discounts or pricing multipliers. The token total includes all billed input, not just newly spoken words. The reviewed documentation does not establish a predictable per-conversation token budget for transcript triggers.
Silence and time spent waiting for backend work still count toward an active Live session. Muting does not close it. OpenAI’s cost guide therefore recommends using final session duration and accounting for backend usage separately. More frequent reactions are not, by themselves, evidence of a cheaper service.
For an initial rollout, choose one reversible reaction or one clearly bounded action. Define what happens when the person changes their mind, how completion is verified and when the session closes. Expand only after evaluating those outcomes—not simply how quickly the assistant starts talking.
Methodology: This AI-assisted guide draws on published OpenAI documentation checked on October 16, 2026, and related RohitAI coverage. Recommendations and arithmetic are analysis, not hands-on results. No deployment, account access or end-to-end speed improvement was tested.