Back Home

推論系統

GPT-Live Voice Agents Can Keep Calls on Standby, Cutting Per-Minute Rates by 60%

The new API is available, and standby rates are listed for existing models too. Developers can reduce voice costs while users wait, but must weigh recovery latency and account separately for backend model and tool usage.

zh-Hant

Voice agents can keep a call connected while users read the screen, look something up, or think, pausing model inference and charging at a lower rate. OpenAI’s GPT-Live documentation now offers a standby setting: after a specified idle threshold, the session pauses and resumes when the user speaks again, without creating a new call. Session documentation

GPT-Live-1.1 now has a model ID that can be used directly in API requests; the official pricing table also lists standby rates for GPT-Live-1 and 1.1. Existing developers can evaluate this cost-saving option without first upgrading models. Model page, pricing table

Keep the connection open, but set an idle threshold

Standby works with WebRTC, the connection method apps such as browsers use to transmit real-time audio. When creating a session, set session.wake_policy.type to speech, then set session.idle_timeout_s to an integer from 5 to 60 seconds. The threshold is based on time without input or output transcripts. If you omit the timeout setting, standby remains enabled after the first wake-up until the session ends. Configuration guide

The microphone audio track must remain enabled during standby. The connection and session ID stay the same, and the conversation history is restored when the session resumes. However, OpenAI warns that the first response may be slower after standby if the user speaks only briefly. Developers must choose a threshold that balances earlier savings against response speed; simply muting the microphone does not put a session into standby. Behavior and limitations

A 60% reduction applies only to time actually spent in standby

Both models cost $0.05 per minute when active and $0.02 per minute in standby, billed by the second. Switching from active to standby saves $0.03 per minute, a 60% reduction for that period. The reduction across an entire call depends on the share of time actually spent in standby. Official pricing table

For example, in a hypothetical 10-minute call with six minutes active and four minutes in standby, voice time costs $0.38. Keeping the call active throughout would cost $0.50, a 24% difference. This example uses list prices; it is not a measured bill. Seconds spent waiting before standby begins still count as active time.

GPT-Live backend model and tool usage are billed separately. The savings above cannot be treated as a reduction in the total bill. If backend costs stay the same, the larger their share of the total, the smaller the share of the total bill saved through voice standby. Billing details

For longer absences, consider closing the session

Standby suits users who are still expected to continue the conversation and need the connection kept open. For long gaps between interactions, the official documentation recommends considering saving the conversation and application task state, closing the voice session, and creating a new one when the user returns. After the session is closed, the app cannot keep listening, so it needs another way to resume, such as a button, push-to-talk, or a wake mechanism. Guidance for handling idle sessions

Standby can preserve continuity through brief pauses; longer absences call for weighing reconnection costs and wait time. These deployment judgments are based on official features and list prices, not independent latency or cost tests.

Sources

  1. Managing GPT-Live sessions | OpenAI API
  2. GPT-Live-1.1 Model | OpenAI API
  3. Pricing | OpenAI API