OpenAI adds GPT-Live-1.1 to its API with $0.02-a-minute standby mode

OpenAI has made GPT-Live-1.1 available in its API and, in the same update, introduced a standby mode that cuts the price of idle voice time from $0.05 to $0.02 per minute. The change appears in the OpenAI API changelog under November 19, 2026, as a feature entry for the `gpt-live-1.1` model and the `v1/live/sessions` endpoint. GPT-Live-1.1 is the full-duplex voice model that OpenAI describes as built "for multilingual conversations", and the company says it improves pronunciation, accent consistency and transcription compared with GPT-Live-1.
Active session time for GPT-Live-1.1 costs $0.05 per minute, the same rate as GPT-Live-1, so developers who switch models pay nothing extra for the newer voice engine. The new standby rate applies to both GPT-Live-1 and GPT-Live-1.1 when they run over WebRTC: once a configurable idle timeout expires, OpenAI pauses inference while keeping the connection open, then resumes when someone speaks again. Prices in this report are as listed on OpenAI's official API pricing page on November 19, 2026, and OpenAI states them per minute of session duration, billed per second.
For any voice application that keeps a session open while the user reads, thinks or walks away from the device, standby reduces the cost of those silent minutes by 60%. Phone calls over SIP and WebSocket connections are outside the scope of the November 19 entry, which names WebRTC sessions only.
··········
GPT-LIVE-1.1 API PRICING BEFORE AND AFTER THE STANDBY CHANGE.
Until this update, OpenAI billed every second of an open GPT-Live session at the active rate. Its cost documentation states that billable time includes periods when the user speaks, when the assistant speaks, when both are silent and when the backend is working, and that muting the microphone does not close the session. Before standby existed, an idle hour of an open session therefore cost $3.00 in voice charges. With standby, the same hour costs $1.20, apart from the few seconds of active time that elapse before the idle timeout triggers.
........
Model | Active session | Standby | Cache | Batch | Effective date |
gpt-live-1 (before) | $0.05 / min | None (idle billed at $0.05) | None listed | None listed | Until Nov 19, 2026 |
gpt-live-1 (after) | $0.05 / min | $0.02 / min | None listed | None listed | Nov 19, 2026 |
gpt-live-1.1 (new in API) | $0.05 / min | $0.02 / min | None listed | None listed | Nov 19, 2026 |
........
OpenAI prices GPT-Live sessions by duration, so per-token rates, cached-input discounts and batch discounts do not apply to the voice session itself. The backend that performs delegated work is billed separately at the normal rates of the configured model and tools. OpenAI's cost documentation gives the formula as billable voice seconds divided by 60 and multiplied by the per-minute voice rate, plus backend costs. Creating a WebRTC session bills 15 seconds of voice duration during initialization, an amount that is credited against duration charges once the session runs.
The rest of OpenAI's live voice price list did not change on November 19. The table below sets GPT-Live against the other per-minute voice products on the same pricing page and against Microsoft's MAI Realtime 1, whose official price is quoted from Microsoft's announcement.
........
Model | Provider | Official price | What it covers |
gpt-live-1.1 | OpenAI | $0.05 / min active, $0.02 / min standby | Full-duplex voice session |
gpt-live-1 | OpenAI | $0.05 / min active, $0.02 / min standby | Full-duplex voice session |
gpt-live-transcribe | OpenAI | $0.017 / min | Live transcription |
gpt-realtime-translate | OpenAI | $0.034 / min | Live translation |
MAI Realtime 1 | Microsoft | From $0.01 / min | Base speech-to-speech, reasoning billed separately |
........
The Microsoft comparison is not like-for-like. MAI Realtime 1 bills reasoning separately at $1 per million input tokens and $5 per million output tokens, it is offered through MAI Labs and a limited-access Microsoft Foundry preview, and Microsoft says preview customers will not be billed until the public preview begins. Our report on Microsoft's MAI Realtime 1 launch covers that model in detail.
··········
HOW GPT-LIVE STANDBY WORKS IN WEBRTC SESSIONS.
OpenAI's guide to managing GPT-Live sessions describes standby as a way to keep a WebRTC connection open during inactivity and resume inference when someone speaks. Developers enable it with two session settings. Setting `session.wake_policy.type` to `speech` makes the session wait for speech before starting inference, and `session.idle_timeout_s` sets how long the model stays active after the last input or output transcript, as a whole number from 5 to 60 seconds. If the timeout is omitted, inference stays active after the first wake until the session ends, which means the session never returns to the cheaper rate.
An optional `session.wake_policy.vad_threshold` controls how much evidence of speech the model needs before waking. It accepts values from 0 to 1, defaults to 0.5, and applies both at the first wake and after each pause. Higher values reduce false wakes from background noise at the cost of ignoring quieter speakers. OpenAI tells developers to keep the microphone track active while the session is on standby, because the model has to hear the user in order to wake.
The following snippet is illustrative and combines the documented field names with the session structure shown in OpenAI's examples. It is not copied from a single official code sample.
{
"session": {
"model": "gpt-live-1.1",
"instructions": "Be concise. Delegate requests needing current information to the backend.",
"wake_policy": { "type": "speech", "vad_threshold": 0.5 },
"idle_timeout_s": 10
}
}Two server events report the state changes. `session.paused` signals that the idle timeout stopped inference and the session entered standby, while the connection and the sideband channel stay open. `session.resumed` signals that inference is ready again, including on the first wake. The connection and session ID stay the same across the pause, so an application does not need to rebuild context or reconnect media. OpenAI warns that after standby the first response to a short utterance may take longer, a latency cost that weighs most on very short exchanges such as yes-or-no confirmations.
··········
WHAT GPT-LIVE-1.1 CHANGES FOR VOICE DEVELOPERS.
OpenAI's model page lists GPT-Live-1.1 with audio and text input, audio and text output, streaming and function calling, and without structured outputs, fine-tuning or predicted outputs. The only snapshot is `gpt-live-1.1`. OpenAI says the model can listen while it speaks, transcribes speech more accurately than GPT-Live-1 and keeps a more consistent accent and voice across languages. The changelog tells developers to use `gpt-live-1.1` in an existing Live session configuration, which makes the upgrade a change of model name for applications already built on GPT-Live-1.
Rate limits are tracked per model. GPT-Live-1.1 and GPT-Live-1 have separate concurrent-session limits, each applying across inference regions, and the GPT-Live-1.1 page points developers to the limits page in their organization settings without publishing tier figures. For GPT-Live-1, OpenAI lists 50 concurrent sessions on the Build tier, 300 on Launch and 500 on Grow, with no access on the free tier.
The model keeps GPT-Live's split between conversation and work. GPT-Live handles listening and speaking, while a backend model looks up information, calls tools and completes tasks, a handoff OpenAI calls delegation. With Responses delegation, OpenAI runs a hosted Responses model and passes context between it and the voice session, and its documentation suggests starting with `gpt-6-luna` and moving to `gpt-6-sol` for more complex backend tasks. With client delegation, the application connects its own model or agent. The November 19 changelog entry adds that developers can describe in plain language when GPT-Live-1.1 should delegate to the backend.
GPT-Live-1.1 reached ChatGPT before the API. OpenAI's system card for the model, published October 15, 2026 on its Deployment Safety Hub, says GPT-Live-1.1 is the default voice model for Plus and higher tiers, including Codex and Dots, while GPT-Live-1.1 Quip is the default for other users. GPT-Live-1.1 Quip does not appear on the API pricing page. ChatGPT users will also lose the older voice modes on December 11, as covered in our report on ChatGPT's switch from Advanced and Standard voice to GPT-Live.
··········
GPT-LIVE STANDBY COST EXAMPLES FOR THREE VOICE WORKLOADS.
The examples below cover voice charges only and exclude backend model and tool costs, which do not change with standby. Each assumes an idle timeout of 10 seconds, so every pause is billed for 10 seconds at the active rate before the session drops to standby.
The first case is a tutoring or companion app in a browser. A 30-minute session contains 12 minutes of conversation and six silent stretches of three minutes each while the learner works through an exercise. Before standby, the full 30 minutes cost $1.50. With standby, 13 minutes are billed at the active rate (12 minutes of talk plus one minute of accumulated timeouts) and 17 minutes at the standby rate, for a total of $0.99, which is 34% less.
The second case is an always-on assistant at a store kiosk or help desk that keeps one WebRTC session open for an eight-hour shift and handles 60 short interactions totaling 45 minutes of speech. Before standby, the shift cost $24.00 in voice charges. With standby, 55 minutes are billed at $0.05 and 425 minutes at $0.02, for $11.25, a reduction of 53%. Over 22 working days the monthly voice bill falls from $528.00 to $247.50 per kiosk.
The third case is a six-minute support call with continuous conversation. It costs $0.30 before and after the change, because there is no idle time to move to the standby rate, and SIP telephony is not among the session types the changelog names.
........
Workload | Assumptions | Before standby | With standby | Change |
Tutoring app (WebRTC) | 30 min open, 12 min talk, 6 pauses of 3 min | $1.50 | $0.99 | -34% |
Kiosk assistant (WebRTC) | 8 h open, 45 min talk, 60 interactions | $24.00 / day | $11.25 / day | -53% |
Kiosk assistant, monthly | 22 working days | $528.00 | $247.50 | -53% |
Support call | 6 min continuous talk | $0.30 | $0.30 | None |
........
Standby favors applications with long, silent stretches inside a single session: tutoring, in-store and hospitality assistants, hands-free cooking or repair helpers, and coding companions that wait while an agent works. Closing the session during inactivity remains cheaper still. In the kiosk example, closing between interactions would leave roughly 55 billed minutes, or $2.75 a day, but OpenAI notes that a closed Live session cannot listen for the user, so the application needs a resume button, push-to-talk control or its own wake trigger. Standby is the middle option for products that have to respond to speech without a tap. The full GPT-Live pricing structure, including telephony partners, is in our guide to the GPT-Live-1 API and its $0.05-per-minute pricing.
··········
GPT-LIVE-1.1 AND STANDBY: CONFIRMED FACTS AND OPEN QUESTIONS.
OpenAI's official documentation confirms the core facts of the update. GPT-Live-1.1 is listed as an active API model at $0.05 per active minute, standby costs $0.02 per minute for both GPT-Live-1 and GPT-Live-1.1 in WebRTC sessions, idle timeouts run from 5 to 60 seconds, and the session keeps its connection and ID while paused. The improvements in pronunciation, accent consistency and transcription are OpenAI's own claims, and the API documentation does not publish benchmark figures comparing the two versions.
Several details remain undocumented. OpenAI has not published concurrent-session limits by tier for GPT-Live-1.1, has not said whether standby will extend to WebSocket or SIP sessions, and does not list GPT-Live-1.1 Quip as an API model. The added latency after a wake is described only qualitatively. For developers already running GPT-Live-1 in browser or mobile apps, the update brings a newer voice model at an unchanged price and a configuration change that can cut idle voice costs by more than half.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




