GPT-Live-1.1 reaches the API, with standby voice at $0.02 a minute
Edited and reviewed by AI Ping Editorial Team
3-line summary
- The November 19 API release brings the voice model to external applications.
- Active voice costs $0.05 a minute and standby $0.02, with backend and tool costs separate.
- WebRTC standby keeps the microphone and connection open, resuming the conversation on speech.
OpenAI has released GPT-Live-1.1 in its API. Its November 19 developer changelog also introduces standby, which pauses inference while keeping a web voice connection open. Active voice costs $0.05 a minute, versus $0.02 on standby.
The model has powered voice in ChatGPT Plus and Pro since October 15. The API release lets developers bring it into their own applications. OpenAI describes improvements in multilingual pronunciation, accent consistency and transcription, without publishing a numerical improvement rate on the model page.
GPT-Live-1.1 uses full-duplex audio: it can listen while speaking, accommodating interruptions and additional information during a reply. The model documentation directs existing Live integrations to select the new model ID. GPT-Live-1 and 1.1 have separate concurrent-session limits, with each model’s limit applying across inference regions.
Standby costs $0.02 a minute
Both models are billed per second at the same rates. Standby’s unit rate is 60% lower. A ten-minute session with four active minutes and six on standby would cost $0.32 for voice duration, versus $0.50 if entirely active. That calculation excludes separately billed backend models and tools.
OpenAI’s cost guide counts silence before the idle timeout, and waits for backend work with inference running, as active time. Default WebRTC startup bills 15 seconds during initialization, credited against duration charges once running. A 90-second active session already includes those 15 seconds; they are not added again. Session creation therefore also matters when an application repeatedly reconnects before a conversation begins.
Standby is configured for WebRTC sessions. Speech can wake inference, and a timeout of five to 60 seconds without an input or output transcript returns it to standby. Omitting the timeout leaves inference active after the first wake until the session ends. Muting a microphone alone does not configure standby.
The microphone track stays active during standby. The connection and session identity persist, and conversation history is restored when speech resumes. OpenAI cautions that the first response to a short utterance after standby may take longer.
Reasoning and tool work go to a backend separately from the spoken conversation. Developers can choose managed delegation to a Responses model or client delegation to their own model, agent or service. In both cases, the application owns permissions, required confirmations, business records and task state. It can review backend results while voice continues. With client delegation, the application chooses the history and task state sent to its backend. It can validate, redact, combine or discard results before the voice model receives them.
Telephone-agent provider ThunderPhone published a same-day comparison of both models on November 19. Approved-reply matching rose from 34% to 43% across 205 moments from real calls; verbatim repetition rose from 56% to 61% across 268 clips. Median time to start speaking fell from 1.3 to 1.0 seconds.
Those figures reflect its phone setup. Transfers, hang-ups and tool timing were excluded. Separate simulated-call examples still showed misspelled names, added phone-number digits and repeated questions. Those examples used invented businesses rather than actual customer interactions.
Sources
- Changelog — November 19, GPT-Live-1.1 — OpenAI
- GPT-Live-1.1 has arrived in the API — OpenAI
- GPT-Live-1.1 model — OpenAI
- Managing GPT-Live sessions — idle sessions — OpenAI
- Pricing — GPT-Live — OpenAI
- Voice session costs and WebRTC initialization — OpenAI
- Delegation and tools in GPT-Live — OpenAI
- GPT-Live-1.1 is faster and follows scripts better, but still slips on spellings and read-backs — ThunderPhone