LATEST
AI Tools

OpenAI GPT-Live-1.1 Is Live: Idle Voice Sessions Cost $0.02 a Minute

Updated November 20, 2026
11 min read
Editor's Note: This article was human-reviewed for accuracy
A headset with a microphone hanging on the side of a laptop in a bright office

Quick Answer & Key Takeaways

  • • OpenAI has opened GPT-Live-1.1, the voice model behind ChatGPT Voice, to developers through its API.
  • • Active voice sessions cost $0.05 a minute, while a new Standby mode drops idle time to $0.02 a minute on WebRTC calls.
  • • Backend reasoning and tools are billed separately, so the voice rate is only part of the total cost.

OpenAI has opened GPT-Live-1.1 to developers. It is the voice model that has powered ChatGPT Voice for Plus and Pro subscribers since October 15, 2026, and it now runs behind any app that wants a voice agent that listens while it talks. The headline number is the price of silence: a new Standby mode bills idle WebRTC sessions at $0.02 a minute instead of the usual $0.05, a 60% cut on the minutes nobody is speaking.

That cut is real, but it is narrow. It applies to idle minutes only, on WebRTC sessions only, and the voice rate is only one line on the bill. The reasoning model and tools behind the voice are charged separately. This article works through the full pricing, what changed from GPT-Live-1, how to build with it, and where the costs hide.

What GPT-Live-1.1 is

GPT-Live is OpenAI's full-duplex voice API. Unlike a push-to-talk assistant, it listens and speaks at the same time, so it can follow pauses, interruptions and changes of pace instead of waiting for a clean turn to end. Sessions run through the /v1/live/sessions endpoint over WebRTC or WebSocket, with SIP support for phone calls.

The important design detail is delegation. GPT-Live-1.1 handles the voice itself, the listening, the speaking, the interruption handling, and hands complex work to a backend model of the developer's choice, through OpenAI's Decisions API or the developer's own agent, without breaking the conversation. That split is why the pricing has two parts: a flat per-minute voice rate, and a separate token bill for whatever the backend does.

GPT-Live-1 launched on September 10, 2026. GPT-Live-1.1 is its successor, and OpenAI announced API access for it on November 19, 2026. It is listed in the API documentation as the recommended model for new Live sessions. If you have used ChatGPT Voice in the last month, you have already talked to this model family.

Each session defaults to a 128,000-token context window. Up to 16,384 tokens of startup instructions are accepted, along with up to 128 prior messages totalling 8,192 tokens. When usage passes 90%, the API starts a summarised replacement voice engine in the background, so a long call does not simply run out of room mid-sentence.

The standby price cut, and what it actually saves

OpenAI's official pricing page lists both GPT-Live models at the same two rates: $0.05 per active minute and $0.02 per standby minute, billed per second. The lower rate applies only to WebRTC sessions while inference is paused. In pounds and Canadian dollars, at November 20, 2026 market rates (1 USD = 0.76 GBP, 1.43 CAD, approximate), the active rate is about 3.8p or 7.2 Canadian cents a minute, and standby is about 1.5p or 2.9 Canadian cents a minute.

Session statePer minute (USD)Per hour (USD)Per minute (GBP, approx.)Per minute (CAD, approx.)
Active$0.05$3.003.8pC$0.072
Standby$0.02$1.201.5pC$0.029

The saving depends entirely on how much of a call is quiet. Take a 30-minute support session with ten minutes of real conversation and twenty minutes of quiet, while the customer looks something up or fills in a form. Kept active the whole time, it costs 30 x $0.05 = $1.50. With the quiet time in standby, it costs 10 x $0.05 + 20 x $0.02 = $0.90. Ten idle minutes on their own cost $0.20 in standby against $0.50 at the active rate.

Scale that to a support line open eight hours a day, with three hours of talking and five hours of waiting across its calls. At the active rate all day, the voice bill is 480 x $0.05 = $24.00. With standby handling the quiet stretches, it is 180 x $0.05 + 300 x $0.02 = $15.00. That is $9.00 a day, or about $198 over a 22-day working month, before the backend bill. A call that is all talking saves nothing at all, which is the first thing to check before assuming the 60% figure applies to your bill.

How standby works

Standby is a configuration on the session, not a separate product. The developer sets an idle timeout between 5 and 60 seconds. After that period of silence, with no input or output transcript, the model pauses inference while the WebRTC connection stays open, along with the session ID and the conversation history. Speech wakes it again, and billing returns to the active rate.

Three states result. Active is when someone is speaking or a response is expected right away. Standby is when the pause is short enough that keeping the microphone and conversation attached is worth $0.02 a minute. Closed is when the user has left, the gap is long, or a backend job will run for minutes without needing voice at all. The policy choice is the developer's: set the timeout too short and the model keeps waking and pausing, set it too long and quiet minutes bill at the active rate.

Two conditions limit the feature. First, standby applies to WebRTC sessions while inference is paused. It is not a blanket discount on every quiet moment or every kind of connection. A WebSocket session or a SIP phone call does not get the standby rate. Second, the backend keeps billing separately throughout. If the voice agent delegates a task to a reasoning model while the caller waits, that work is charged at the backend model's token rates, even while the voice side sits in standby.

What changed from GPT-Live-1

OpenAI says GPT-Live-1.1 speaks with more natural and consistent accents across languages, understands precise details such as letters and numbers better, follows instructions more reliably, and completes delegated tasks more consistently. Those are OpenAI's own claims, from its announcement, and no independent evaluation of the API models had been published at launch. The updated system card, first published on October 15, 2026, gained an addition on November 19, 2026 with evaluations of the API voice models, run with backend delegation disabled.

The practical differences for a developer are the model recommendation and the standby option. The Live getting-started guide now recommends GPT-Live-1.1 for new sessions, and standby is available on both GPT-Live-1 and GPT-Live-1.1, so existing GPT-Live-1 apps get the cheaper idle rate without switching models. The active rate did not change. Anyone reporting this as a price cut on voice calls is overstating it: the per-minute price of talking is exactly what it was.

The model takes text and audio as input and output. Images and video are not supported, which matters for any agent that needs to see a document or a photo during a call. Streaming and function calling are supported. Structured outputs and fine-tuning are not, so teams that need strict JSON schemas or a tuned voice will need to handle that in the backend.

How to build with it

The quickstart in OpenAI's Live guide keeps the moving parts small. You need a microphone, a browser page served over HTTPS or localhost, and a trusted server holding an OpenAI project API key. The key stays on the server, never in the browser.

  1. On the server, create a session and set the model to gpt-live-1.1, with a short conversation prompt that tells GPT-Live when to ask the backend for help.
  2. Set the standby idle timeout in the session configuration, between 5 and 60 seconds, if you want quiet time billed at the lower rate.
  3. Have the browser send its WebRTC SDP offer to your server, which exchanges it for an answer from the Live API and returns it to the browser.
  4. Wait for the session.started event, then speak and listen for a reply. The browser handles the microphone, audio playback and event channel.
  5. Wire the delegation target, a Decisions API agent or your own backend, so the voice model can hand off work that needs tools or reasoning.

For phone use, the same sessions endpoint supports inbound SIP with a signed webhook and outbound SIP through the customer's trunk, with call controls for accept, reject, refer, hangup and fork, plus stored sessions for later review. Ringing is limited to three minutes and a connected call to two hours.

Voice choice is wider than a single default. The default voice is marin, with 21 additional voices listed in the session guide, covering English, Portuguese, German, French, Hindi, Japanese and Korean. Prompt-generated voices can be created, up to 100 per organisation for GPT-Live only. Voices cloned from recordings are limited to eligible customers, at most 20 per organisation.

GPT-Live-1.1 vs the alternatives

Voice agents are priced in different ways, so a flat per-minute rate is not automatically cheap. The table compares the four voice options our readers ask about most, using published rates where they exist.

Voice APIVoice billingBackend billingTransportsAccess
OpenAI GPT-Live-1.1$0.05/min active, $0.02/min standby (WebRTC only), per secondSeparate, per tokenWebRTC, WebSocket, SIPLive in API since Nov 19, 2026
OpenAI GPT-Live-1$0.05/min active, $0.02/min standby (WebRTC only), per secondSeparate, per tokenWebRTC, WebSocket, SIPLive since Sep 10, 2026
Mistral Voxtral 3$0.04/min with Mistral Small 4, $0.08/min with your own LLMIncluded in the per-minute rateAPI, SDK, self-deployLive since Oct 2026
ElevenLabs Eleven RealtimeNo published per-minute rate at launchPer-minute-plus-tokens API billing, per launch coverageAPI, WebRTCLive since Oct 2026
Block DessaNo price publishedNot disclosedNot disclosedWaitlist only

The honest comparison is with Voxtral 3, which we reviewed in October. Voxtral charges $0.04 a minute with its own small model, a fifth less than GPT-Live's active rate, and that price includes the reasoning. GPT-Live looks cheaper only when the call is mostly quiet, because Voxtral has no standby tier, and only before the backend bill. A GPT-Live call that delegates heavily to GPT-6.2 Astra at $8 per million input tokens and $40 per million output tokens can pass Voxtral's flat rate quickly. The cheaper backend choice is GPT-6.2 Sol at $2 and $10 per million tokens, and the gap between the two backends is a real design decision, not a rounding error.

ElevenLabs is harder to compare because it published no per-minute price at launch, which we flagged in our earlier coverage. Its billing mixes per-minute and per-token charges, so a buyer cannot compare it to OpenAI's flat rate without running their own workload through both. Block's Dessa remains waitlist-only with no price at all, so it is a product to watch rather than a price to compare.

The catches

Five limits decide whether GPT-Live-1.1 fits your use case. First, standby is WebRTC-only. If your calls arrive over SIP or WebSocket, every minute bills at $0.05, and the 60% figure never applies. Second, the backend bill is uncapped by the voice rate. A voice agent that reasons hard on every turn can cost several times the voice line, and OpenAI bills the two parts separately, so the invoice needs both lines modelled.

Third, there is no self-hosted option and no Free tier. Concurrent sessions are capped at 50 on the Build tier, 300 on Launch and 500 on Grow, which is enough for a pilot and a constraint for a large contact centre. Fourth, calls are capped at two hours and ringing at three minutes, so long-running sessions need reconnection logic. Fifth, the model cannot see images or video, cannot return structured outputs, and cannot be fine-tuned, so document-heavy or form-filling agents need the backend to do that work.

The common mistake is to budget from the headline. A team that reads "60% cheaper" and multiplies its whole voice bill by 0.4 will under-budget. The correct estimate is active minutes times $0.05, plus eligible standby minutes times $0.02, plus backend tokens at the backend model's rates, plus telephony and infrastructure. Standby only helps if your calls actually pause, and the pause has to happen on WebRTC.

Who should use it

GPT-Live-1.1 fits support and sales calls where the agent and the caller take turns and the caller regularly goes quiet to check something, fill in a form, or wait for a human colleague. Those are the calls where standby pays. It also fits teams already on OpenAI's stack, since delegation to the Decisions API keeps the voice and the reasoning in one vendor's billing.

It fits less well for always-on monitoring, phone-first deployments, or calls that run for hours, where the WebRTC condition, the two-hour cap and the separate backend bill all work against it. Teams comparing it with Voxtral should price one real week of their own calls through both, including the backend, rather than comparing headline per-minute rates. The voice rate is the smaller number in most real deployments.

Frequently asked questions

How much does GPT-Live-1.1 cost per minute?

Active voice sessions cost $0.05 a minute, billed per second, and standby costs $0.02 a minute. That is about 3.8p or 7.2 Canadian cents active, and 1.5p or 2.9 Canadian cents on standby, at November 20, 2026 rates. Backend model and tool usage is billed separately, on top of the voice rate.

What is Standby mode in GPT-Live?

Standby is a session setting that pauses the model after a configurable idle timeout of 5 to 60 seconds of silence. The WebRTC connection, session ID and conversation history stay open, and speech wakes the model again. While paused, the session bills at $0.02 a minute instead of $0.05.

Does standby work on WebSocket and phone calls?

No. OpenAI's pricing and model pages state that the lower standby rate applies to WebRTC sessions while inference is paused. WebSocket sessions and SIP phone calls bill at the active rate for their full duration, including quiet stretches.

What backend model does GPT-Live-1.1 use?

GPT-Live-1.1 does the voice work itself and delegates reasoning and tools to a backend the developer chooses, through OpenAI's Decisions API or the developer's own agent. The backend is billed at that model's own token rates, for example GPT-6.2 Sol at $2 per million input tokens and $10 per million output tokens, or Astra at $8 and $40.

Can I self-host GPT-Live-1.1?

No. GPT-Live is cloud-hosted only, with no self-hosted or on-premises option. The Free tier is also excluded, and concurrent sessions are capped by tier, at 50 on Build, 300 on Launch and 500 on Grow.

Is GPT-Live-1.1 the same as ChatGPT Voice?

It is the same model family. GPT-Live-1.1 has powered voice in ChatGPT Plus and Pro since October 15, 2026. The November 19 announcement opened that model to developers through the API, so the news is API access and the standby option, not a new model.

Sources

11 min read
25 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more