OpenAI just released GPT-Live-1.1 in the API, along with a new standby mode that makes long-running voice AI conversations significantly cheaper. And honestly, the standby feature might be the more important update. Until now, keeping a voice AI session open meant paying for the entire session, including those awkward moments when nobody was speaking. Now your AI assistant can effectively go to sleep without hanging up. When the conversation becomes inactive, inference pauses. The WebRTC connection stays open, and when you start speaking again, the same session resumes with its conversation history intact. And the price drops from $0.05 to $0.02 per minute during standby. WHAT'S NEW IN GPT-LIVE-1.1? GPT-Live-1.1 is OpenAI's updated full-duplex voice model for developers. It can listen while speaking, handle interruptions, and delegate complex tasks to backend models and tools. Compared with GPT-Live-1, OpenAI claims several improvements: • More accurate speech transcription. • Better recognition of details such as letters and numbers. • More consistent accents and voice characteristics when switching between languages. • Improved instruction following and task completion. • More reliable multilingual conversations. However, OpenAI hasn't published specific benchmark numbers showing how much better GPT-Live-1.1 performs. So these are company-reported improvements, not independently verified performance gains. HOW DOES STANDBY WORK? The idea is simple. Imagine you're talking to an AI assistant while working. You ask a question. The assistant responds. Then you spend a few minutes reading, writing, or thinking. Normally, the voice session would keep running at the active rate. With standby enabled, OpenAI temporarily pauses inference after a configurable period of inactivity. The connection remains open. The session ID remains unchanged. The conversation history is retained. And the model resumes when you start speaking again. No need to reconnect or repeat the previous conversation. HOW MUCH DOES IT COST? Both GPT-Live-1 and GPT-Live-1.1 use the same pricing: ACTIVE SESSION: $0.05 per minute STANDBY: $0.02 per minute Billing is calculated per second. That's a 60% reduction for time actually spent in standby. For example, imagine a 30-minute voice session with five minutes of conversation and five separate five-minute pauses. Without standby: 30 minutes × $0.05 = $1.50 With a 15-second idle timeout: 6.25 active minutes + 23.75 standby minutes = approximately $0.79 That's almost half the voice session cost. Actual savings depend on how frequently the user speaks and how long the pauses last. Backend models and tools are billed separately. HOW TO ENABLE STANDBY Standby is available for GPT-Live-1 and GPT-Live-1.1, but only through WebRTC sessions. When creating a session, configure: session.wake_policy.type = "speech" session.idle_timeout_s = 15 The idle timeout accepts a value between 5 and 60 seconds. For example, with a 15-second timeout, inference pauses after 15 seconds without incoming or outgoing transcribed speech. When speech is detected again, the session wakes up. Developers can also monitor the session using two events: session.paused session.resumed There's an optional voice activity detection threshold that determines how confidently the system must detect speech before waking. THE IMPORTANT CATCH Muting your microphone doesn't activate standby. In fact, OpenAI explicitly recommends keeping the microphone track active so the server can detect when you start speaking again. If the microphone is muted, the session remains active and continues billing at $0.05 per minute. Standby is also not available for WebSocket or SIP voice sessions. And there's a trade-off: the first response after waking from standby may take longer, particularly after a short utterance. So standby is useful, but not free, instant, or universally supported. OFFICIAL SOURCES GPT-Live-1.1 model documentation: https://developers.openai.com/api/docs/models/gpt-live-1.1 Live session management and standby: https://developers.openai.com/api/docs/guides/live-conversations GPT-Live cost optimization: https://developers.openai.com/api/docs/guides/voice-latency-cost OpenAI API pricing: https://developers.openai.com/api/docs/pricing Original announcement: https://x.com/OpenAIDevs/status/2123506733524136446 Cost breakdown and examples: https://tproger.ru/news/openai-otkryla-gpt-live-1-1-v-api-i-udewevila-pauzy-v-razgovore Would you keep a voice AI assistant connected throughout your workday if standby made it cheaper? Or do you prefer assistants that only activate when explicitly called? #GPTLive #OpenAIAPI #VoiceAI