The model ChatGPT users already had

GPT-Live-1.1 is not a new model in the usual sense. ChatGPT Plus and Pro users have been talking to it since October 15. What changed on November 19 is who else can build on it: the model is now in the API.

The architecture is the part worth understanding. GPT-Live handles the conversation itself, listening and speaking at once, and can delegate deeper reasoning and actions to the models and tools it is paired with — OpenAI's own wording for the family. Developers choose the models, tools and agent harness behind the conversation. The voice layer and the thinking layer are separate products with separate bills, and that split runs through everything about the pricing.

Silence gets its own price

Voice sessions spend much of their life waiting: a caller reading a confirmation code, a customer looking for an order number, a patient checking a date. Until now, an open GPT-Live session cost the active rate through those pauses. Standby changes the unit economics of waiting. After the timeout the developer sets, between 5 and 60 seconds, inference pauses and the session rate falls to $0.02 a minute until someone speaks again.

The saving is real but bounded. It applies to the voice layer, which at $0.05 a minute was never the whole cost of a call. A session that delegates hard work to a frontier backend model will still pay frontier prices for that work, and Standby does nothing to that line. Read the announcement as a 60% cut to one component of the bill, not to the bill.

What OpenAI did not quantify

The improvement claims in the announcement come without numbers. More consistent accents, better handling of spelled letters and numbers, more reliable task completion: each is plausible for a point release, and none is given a score, a test set or a comparison table in the announcement. The earlier GPT-Live-1 launch at least published evaluation results, including a 30-percentage-point gain on Full Duplex Bench for GPT-Live-1 over GPT-Realtime-2.1. For 1.1, developers are asked to take the upgrades on trust and test them on their own call traffic.

That is a workable arrangement for teams already running voice agents, who can measure the model on their own traffic. It is a weaker basis for anyone choosing a voice stack from scratch, where a point release with no published numbers has to be evaluated from zero.

The practical read

If you already run GPT-Live sessions with long pauses, Standby is a direct saving: set the timeout, keep the call open, pay less for silence. If you are budgeting a new voice product, build the estimate in three lines from the start — active voice minutes, Standby voice minutes, and backend model and tool usage — because OpenAI prices them separately. The headline rate of $0.05 a minute, or $0.02 idle, covers only the first two.

Prices as published by OpenAI and reported on November 19, 2026; check the API pricing page before committing.

Sources

  1. [1] OpenAI, "Build more natural voice experiences with GPT-Live-1 in the API" — family launch post; $0.05/min voice-layer pricing, delegation architecture, Full Duplex Bench resultsRead source
  2. [2] OpenAI developer announcement of GPT-Live-1.1 in the API, November 19, 2026, as relayed by AlphaSignal and Superpower Daily (label only, no URL recorded)
  3. [3] AlphaSignal, "OpenAI Ships GPT-Live-1.1 to Developers With 60% Cheaper Idle Sessions," November 19, 2026Read source
  4. [4] Vercel changelog, "GPT-Live 1.1 now available on AI Gateway," November 19, 2026Read source