
GPT-Live-1.1 Reaches the API: Standby Now Costs $0.02
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 358 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 139 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 80 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 111 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Five weeks after GPT-Live-1.1 replaced its predecessor inside ChatGPT Voice, it became callable. The model went live on the OpenAI API on November 19, 2026 under the ID gpt-live-1.1, reachable only through v1/live/sessions. The changelog entry is three sentences long and mentions pronunciation, accent consistency and transcription. The pricing page is where the actual news is: alongside the familiar $0.05 per active voice minute there is now a second line, $0.02 per minute, for a session that is connected but not thinking. Silence, which used to cost the same as speech, is now 60% cheaper.
That second line is the part worth understanding, because it changes the shape of a cost model rather than the size of one number. This is a release note about a model swap married to a billing change, and the billing change is the one that will show up in your invoice.
What shipped on November 19
The model page describes GPT-Live-1.1 as a full-duplex voice model for multilingual conversations, and the only capability statement attached to the version bump is comparative and unquantified: compared with GPT-Live-1, it "transcribes speech more accurately and maintains a more consistent accent and voice across languages." There is no benchmark table, no word error rate, no list of which languages improved by how much. That is worth saying plainly, because a version number invites the assumption of a measured improvement, and OpenAI has published the claim rather than the measurement.
Structurally nothing else moved. Input and output modalities are audio and text. The supported feature list is streaming and function calling. Images and video are not inputs. Chat Completions, Responses, Realtime, transcription, translation, speech generation, batch, fine-tuning, embeddings and moderation are all listed as not supported — the Live sessions endpoint is the only door in, which is the same arrangement GPT-Live-1 has had since it went generally available on September 10, 2026.
What did accumulate around the model is voice inventory. The Live documentation now lists 21 named voices spanning English in Australian, British, Irish, North American, Filipino and Southern U.S. accents, plus Brazilian Portuguese, German, French, Hindi, Japanese and Korean. Nine of those — Nuri, Juni, Flitz, Brise, Sillage, Nira, Harema, Shida and Noeul — were added on October 16, 2026, alongside prompt-to-voice becoming generally available to all API customers. Prompt-generated voices are free to create and billed at normal session rates when used, and they work with GPT-Live but not with the standalone text-to-speech endpoint.
Standby is the real change

A voice session that stays open while nobody is talking is the normal condition of a phone agent, an in-car assistant, or a support line with a customer on hold. Before this week you paid the active rate for that silence. Now you can configure the session to stop inferring after a set period without speech and resume when someone talks, and the paused time bills at $0.02 a minute.
The configuration is small and specific:

• Wake policy — set session.wake_policy.type to speech so the session waits for speech before it starts inferring, rather than starting on connection.
• Idle timeout — set session.idle_timeout_s to a whole number between 5 and 60. That is how many seconds without an input or output transcript elapse before the session returns to standby. Omit it and inference stays active after the first wake until the session ends, which is the old behaviour with a new name.
• Sensitivity — session.wake_policy.vad_threshold takes a number from 0 to 1 and defaults to 0.5. Higher values require stronger evidence of speech before the session wakes, which is the knob you turn when a noisy room keeps waking it.
• Events — session.paused fires when the idle timeout stops inference; session.resumed fires when inference is ready again, including on the first wake.
Two operational details matter more than the parameter list. The microphone track stays active while the session is on standby, and the connection and session ID stay the same — the session resumes and restores conversation history as the user speaks, so you are not re-authenticating a caller or rebuilding context. And standby is a WebRTC feature: the documentation describes it for WebRTC sessions, and the idle-session guidance for other transports is to close the session and start another one. If you are on WebSockets or SIP, the cheaper rate is not the lever you have.
The cost is not zero, either. The docs are explicit that after standby, the first response to a short utterance may take longer, because the model is restarting inference rather than continuing it. You are trading latency on the first turn after silence for 60% off the silence itself.
What a minute of silence costs
Both rates bill per second, so the arithmetic is continuous rather than rounded to the minute. OpenAI's own worked example is a call that spends 60 seconds active and 30 seconds on standby: one minute at $0.05 plus half a minute at $0.02 comes to $0.06 for 90 seconds of connected time. Had all 90 seconds been active, the same call would have cost $0.075.
Scale that to something that looks like a deployment. An eight-hour connected shift at 30% standby is 5.6 active hours at $3.00 and 2.4 standby hours at $1.20 — $19.68 against $24.00, a saving of $4.32 per shift, or about 18%. Push standby to half the shift and the same day costs $16.80, a 30% cut. These are voice-line figures only. Backend model and tool usage is billed separately and, in a real deployment, is frequently the larger number.

One charge survives the optimisation: creating a WebRTC session with POST /v1/live/sessions bills 15 seconds of voice duration while the session initialises. That amount is credited against duration charges once the session starts running, so it is not a surcharge on a session that does work — but it is a real cost for an application that opens sessions speculatively, and it makes "close the session and start another" less free than it sounds for short interactions.
The backend is a separate bill
GPT-Live is a two-part system, and the version bump did not change that. The voice model manages the spoken conversation; reasoning and tool use are delegated across an asynchronous boundary to a backend model or agent you configure. With Responses delegation, GPT-Live calls the Responses model you name and returns its results to the live conversation. With client delegation, your application prepares the context, runs the workflow, and sends results back — which is the mode to pick when you need custom routing between models, fallbacks, or budgets across backend steps.
OpenAI's documentation recommends starting with gpt-6-luna and moving to gpt-6-sol for more complex backend work, and both are on the standard token meter: $0.10 and $0.50 per million tokens for Luna, $2.00 and $10.00 for Sol, input and output respectively. Over a long call that backend spend can exceed the voice line entirely, which is why the standby discount should be read as a fix to one line of the bill rather than to the bill.
This is also where a router earns its place. OrcaRouter does not serve GPT-Live-1.1 — the Live sessions endpoint is OpenAI's own, and the catalogue does not carry a full-duplex voice model. What it does serve is the half of the system that does the thinking: gpt-6-luna, gpt-6-sol and the rest of the GPT-6 family sit behind one key at provider list price with no markup, so the token side of a Live deployment can be routed, failed over, and priced the same way it would be anywhere else. If you are evaluating which backend makes a $0.05 voice minute worth paying for, that comparison is cheap to run through one endpoint.
The open question
Nobody outside OpenAI has published a measurement of GPT-Live-1.1 yet. The Artificial Analysis speech-to-speech index, read on November 20, 2026, carries GPT-Live-1 twice — 83.2 on the index with an Astra-class backend at medium reasoning effort, 80.1 with a Sol-class backend at low — and does not carry 1.1 at all. That is not a criticism of the release; the model is a day old on the API and independent evaluation takes longer than that. It is a reason to treat the accent and transcription claims as vendor-reported until they are not, and a reason to test the specific thing you are buying: if your product runs Spanish and Japanese, the improvement in accent consistency is either audible in your own recordings or it is not, and no third party can tell you yet.
The practical read for this week is narrow. If you already run GPT-Live-1, the upgrade is a model ID change with no price movement and no API change, and the standby rate is available to both models — the changelog entry covers GPT-Live-1 and GPT-Live-1.1 WebRTC sessions together. Set an idle timeout before you set anything else, because that is the only item in the release that changes what you pay.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
