Article

GPT-Live standby costs and when to close an idle session

For teams building voice assistants, compare active and standby charges, count idle time correctly and decide when reconnecting is worth it.

Editorial illustration for GPT-Live standby costs and when to close an idle session: a geometric block represents a model release. Not documentary evidence.

OpenAI’s November 19, 2026 update adds standby for GPT-Live-1 and GPT-Live-1.1 voice apps using WebRTC live-audio connections. For teams handling intermittent conversations, use paid standby when keeping the connection matters; close longer gaps if the app can reconnect and restore the conversation.

Both models cost US$0.05 per active minute and $0.02 per standby minute, billed per second. Staying active avoids the standby wake-up step, but you keep paying the higher rate during quiet periods. You do not need to upgrade to GPT-Live-1.1 just to use standby.

What changes when a session enters standby?

Standby pauses model processing while retaining the connection and session ID. Speech resumes the session with its history, but the microphone track must remain active. A short utterance may receive a slower first reply; OpenAI gives no numerical wake-up delay in its standby documentation.

At session creation, set session.wake_policy.type to speech. Set session.idle_timeout_s to an integer from 5 to 60: inference pauses after that many seconds without an input or output transcript. Omit the timeout and inference stays active after its first wake.

Product implication: show an accurate listening state. Standby should not look like microphone-off. Muting alone does not stop active billing, and waiting for backend work also counts while inference runs.

How much does a ten-minute session cost?

Use billed time in each state, not simply the time someone spends talking. These illustrative calculations use the November 19 list rates and cover voice duration only:

  • Continuously active: 10 minutes × $0.05 = $0.50.

  • Two billed active minutes plus eight billed standby minutes: (2 × $0.05) + (8 × $0.02) = $0.26. That is $0.24, or 48%, below the continuously active example.

Eight quiet minutes are not necessarily eight standby minutes. In a separate hypothetical ten-minute period, assume two active minutes followed by eight quiet minutes and a 30-second timeout. The first half-minute of quiet still costs the active rate, giving 2.5 active minutes and 7.5 standby minutes: (2.5 × $0.05) + (7.5 × $0.02) = $0.275, or 27.5 cents.

That second calculation assumes the timeout starts at the two-minute mark, with no intervening transcripts, wakes, playback or additional billed duration. The active-time billing rule is why the saving becomes 45%, not 48%.

Neither example is a full application quote or measured saving. They assume no discounts and exclude additional starts or retries, taxes, hosting, external services and separately billed backend models and tools. For a different voice-stack budgeting exercise, see PolyAI Dialog-RSN-1 pilot costs on AWS; its host-hour charges are not directly comparable with these session-minute rates.

When is closing worth the interruption?

Closing stops voice charges, so consider it when the user leaves or does not need speech while a longer task runs. With client delegation, application-managed backend work can continue while voice is closed, and its costs still need counting.

An illustrative cost rule is: close when the standby charges you would avoid exceed the genuinely additional cost of restarting. For a gap of G minutes, the avoided standby charge is $0.02 × G. Reconnection delay and conversation continuity are separate trade-offs; the documentation does not establish a universal cutoff.

Do not manufacture a fixed restart penalty. Under default WebRTC startup behavior, 15 seconds of voice duration is billed during initialization, then credited against running-session duration. It is not another 15 seconds added to every running session. This wording does not settle every failed-start, very-short-session or speech-wake case.

A closed session cannot hear a returning user. Provide a resume control or app-managed wake trigger, preserve context and task state, and finish playback and required work before closing. Follow the close-and-restart procedure; a gap between transcript events alone is not proof of silence.

Reconnecting must not repeat a completed booking or other business action. OpenAI’s delegation guidance leaves permissions, confirmations and task records with the application. Our guide to when GPT-Live assistants should act covers that separate decision: restoring a conversation is not authorization to run its actions again.

What should your team verify before rollout?

Confirm GPT-Live project access and microphone permission and check your account’s concurrent-session limits. The two models have separate limits shared across inference regions. The cited docs do not establish that standby frees a session slot, a maximum standby duration or universal regional eligibility. This standby configuration is for WebRTC, not a rule to copy into Realtime, WebSocket or SIP integrations.

Evaluate short replies, background noise, reading pauses and reconnection with your own workload before choosing a cutoff. Compare completed tasks and response delays alongside the bill. Track active and standby transitions for estimates, and reconcile against provider billing records. Usage updates are cumulative: retain the final total when the session closes rather than adding successive snapshots.

The useful policy is conditional: stay active when avoiding wake-up delay is worth the cost; use standby for connected pauses; close when the person is away and a deliberate return is acceptable. Choose the boundary from your application’s needs, not from an assumed free idle state.