By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
2 min read
Key Takeaways
OpenAI added GPT-Live-1.1 to its paid API on November 19. For opt-in WebRTC standby, OpenAI lists $0.02 per minute instead of the $0.05 active rate on both GPT-Live versions. OpenAI sells the API; its reported voice-quality improvements have not been independently tested here.
What OpenAI changed in GPT-Live-1.1
OpenAI's API changelog entry for November 19 adds gpt-live-1.1 to the Live sessions endpoint, v1/live/sessions. GPT-Live is full-duplex: the model can listen while it speaks, so a caller can interrupt instead of waiting for a turn to end. Against GPT-Live-1, OpenAI says 1.1 transcribes speech more accurately and holds a more consistent accent and voice across languages, with better pronunciation. Those are OpenAI's own comparisons; the voice-quality improvements have not been independently tested here.
OpenAI says GPT-Live-1.1 began rolling out in ChatGPT Plus and Pro voice on October 15. Developers can now select it when creating a Live API session. Context events are unchanged, but OpenAI's migration guide uses plain-language instructions for backend delegation instead of the special tokens used by GPT-Live-1.
Active vs. standby: the new meter
OpenAI lists per-second billing at effective rates of $0.05 per minute active and $0.02 per minute in WebRTC standby. That makes the standby session rate 60% lower than the active rate. An hour held open entirely in standby costs $1.20 instead of $3.00 at the active rate. Backend model and tool usage is billed separately in both states, so standby cuts the session charge, not the whole bill.
Developers turn it on at session creation: set the wake policy to wait for speech, and optionally set a 5–60 second idle timeout. The timeout counts seconds with no input or output transcript; when it expires, inference pauses but the WebRTC connection, the session ID and the conversation history all stay in place, and the session resumes when someone speaks. Standby works for GPT-Live-1 sessions too, not just 1.1.
Where standby applies
Standby is WebRTC only — the lower rate does not apply while inference is running, and the standby option does not apply to WebSocket or SIP sessions. Leave the idle timeout out and a session stays at the active rate after its first wake until it ends. The microphone track has to stay live while a session sits in standby, and OpenAI warns the first response after standby, especially to a short utterance, may take longer than usual.
The model's limits are unchanged elsewhere: structured outputs, fine-tuning and predicted outputs are not supported, while streaming and function calling are. GPT-Live-1 and GPT-Live-1.1 also carry separate concurrent-session limits, so a fleet running both models is capped on each one independently.
What changes the bill
Savings on the voice session depend on how long it spends in standby, not simply how long the connection remains open. The listed lower rate applies to both GPT-Live versions. Backend model and tool charges remain separate.