OpenAI GPT-Live-1.1: Voice Agents That Cost 60% Less When Nobody Is Talking

Talking to an AI used to mean waiting your turn. You spoke, the model thought, then it answered — and if you tried to interrupt, you were out of luck. OpenAI’s GPT-Live line changed that with full-duplex voice: a model that listens while it speaks, so a caller can cut in mid-sentence the way people do in a real conversation. Now the line has a new version. GPT-Live-1.1 is available in OpenAI’s API, listed in the API changelog on November 19, 2026, and its headline feature is not about how it talks — it is about what it costs when nobody is talking at all.

A new standby mode drops the price of idle time in a voice session from $0.05 to $0.02 a minute, a 60% cut. For anyone building a voice agent that stays on the line while a customer reads, thinks or walks away from the device, that changes the economics of the whole product. Here is what GPT-Live-1.1 is, how standby works, and when it makes sense to build on it.

What Is GPT-Live-1.1?

GPT-Live-1.1 is a speech-in, speech-out model built for real-time voice conversations. OpenAI’s model page describes it as a “full-duplex voice model for multilingual conversations,” and the full-duplex part is the point: the model hears and speaks in a single stream, so interruptions work naturally instead of breaking the session. It is aimed squarely at developers building voice agents — customer-support lines, phone agents, voice assistants inside apps — rather than at chat users. That is a different job from ChatGPT’s live dictation, which types out what you say; GPT-Live-1.1 holds the conversation itself.

The architecture splits the work in two. GPT-Live handles the voice — listening, speaking and keeping the conversation flowing — while the reasoning and tool calls are handed to a backend model you choose. You reach it through one endpoint, v1/live/sessions, and billing works by duration rather than by tokens: the voice session is charged per minute, and the backend model and tool usage are billed separately at their normal rates. That separation matters when you estimate costs, because the headline price is only part of the bill.

The 1.1 update also improves the voice itself. Compared with GPT-Live-1, the new version transcribes speech more accurately and holds a steadier accent and pronunciation across languages, according to OpenAI’s own comparison. It keeps the same endpoint, pricing and feature set as its predecessor, so switching models costs nothing extra — and the things GPT-Live-1 could not do, such as structured outputs, fine-tuning and predicted outputs, are still unsupported.

Standby Mode: The 60% Cheaper Idle Minute

The problem standby solves is simple, and anyone who has run a voice agent has felt it. A voice session used to bill every second at the same $0.05-a-minute active rate, whether someone was talking or the line was silent. A customer on hold, a user reading a document aloud to themselves, a caller who walked away from the phone — all of it cost the same as an actual conversation. For an agent that keeps a session open for an hour, that adds up fast.

Standby changes the maths. Once a configurable idle timeout expires, OpenAI pauses inference while keeping the WebRTC connection open, then resumes when someone speaks again. The idle minutes bill at $0.02 instead of $0.05 — a 60% cut. An hour of pure idle time drops from $3.00 to $1.20, and because sessions are billed per second rather than rounded up to whole minutes, you only pay for the silence you actually use. The connection and session ID stay the same through standby, and the conversation history comes back when the session resumes, so the caller never notices the pause.

There is one important limit: standby is WebRTC-only. It applies to browser-based voice sessions, not to phone calls over SIP or to WebSocket connections. It also applies to both GPT-Live-1 and GPT-Live-1.1, so existing deployments get the cheaper idle rate too — but the new model is the one OpenAI is steering developers toward.

How to Turn Standby On

Standby is not a dashboard toggle. You configure it when you create the session, with two settings doing the work. Setting session.wake_policy.type to “speech” tells the session to wait for voice activity before starting inference, and session.idle_timeout_s sets how long the model stays active after the last input or output transcript — a whole number from 5 to 60 seconds. An optional session.wake_policy.vad_threshold, from 0 to 1 with a default of 0.5, controls how much evidence of speech is needed to wake the session; raise it if background noise keeps waking it early.

Your app can watch the session state through the data channel: a session.paused event means standby has started, and session.resumed means inference is running again, so your interface can show the right state to the user. One practical detail matters: the microphone track has to stay open during standby, because the idle timer measures silence rather than a muted mic. Set the timeout to match your use case — a short timeout for a fast-paced support line, a longer one for a companion agent where pauses are part of the conversation.

For a step-by-step walkthrough of the new endpoint and settings, see Apidog’s GPT-Live-1.1 guide.

A developer wearing headphones at a home office desk, with code and a glowing voice waveform on a laptop screen

GPT-Live vs GPT-Realtime: Which Voice API Fits?

OpenAI now has two ways to build a voice product, and picking the wrong one is an expensive mistake. GPT-Live bills by the minute: $0.05 active, $0.02 on standby, with the backend billed separately. GPT-Realtime bills by the audio token — $32 per million input audio tokens and $64 per million output tokens — which makes its cost harder to predict and, for long sessions, often higher.

The simple rule is that GPT-Live fits always-on agents with long, unpredictable sessions, because its per-minute pricing is easy to forecast and standby makes the quiet stretches cheap. GPT-Realtime fits shorter, denser interactions where you want fine-grained control over the model itself. If your agent spends most of its time waiting — a receptionist, a tutor, a voice companion — Live with standby is the cheaper architecture. If it spends most of its time talking, compare the token maths before you commit.

What It Means for Voice Agents

The timing is no accident. Voice agents are moving from demos to real products: Microsoft Perception now ships AI agents that attack, defend and fix networks, Google’s Gemini 4 Neon is pushing frontier-level reasoning into cheaper, faster tiers, and even open models like Inkling 2 Mini are built to run voice-capable assistants on a laptop. The bottleneck is no longer whether a model can hold a conversation — it is whether the business case survives the phone bill. Standby attacks that bottleneck directly.

That makes three kinds of product meaningfully cheaper to run. Customer-support lines that keep callers on hold, voice companions that sit quietly between exchanges, and in-app assistants that wait while a user reads or types all spend most of their session time idle. A 60% cut on those minutes is the difference between a feature that pays for itself and one that gets switched off after the first invoice. And because the standby rate applies to GPT-Live-1 as well, teams already running voice agents get the saving without changing a line of model code — only the two session settings.

There are still things to watch. The backend model is billed separately, so a session that talks to a frontier reasoning model is not a $0.05-a-minute product. Standby covers WebRTC only, so phone-based agents on SIP do not get the discount yet. And the voice model still cannot do structured outputs or fine-tuning, so agents that need strict JSON or a custom-trained voice will have to look elsewhere. But for the broad middle of voice agents — the ones that listen, answer and wait — GPT-Live-1.1 with standby is the cheapest way OpenAI has offered to keep an AI on the line.

A smart speaker and a smartphone on a wooden table with glowing teal and orange sound waves flowing between them

GPT-Live-1.1 is a small update with a large consequence. The voice itself is better — steadier accents, more accurate transcription — but the real story is standby mode, which cuts the cost of silence by 60% and makes always-on voice agents affordable to run. At $0.02 a minute for idle WebRTC time, billed per second, a voice agent can stay on the line for an hour for $1.20 plus the cost of the moments it actually speaks. That is the price point at which voice stops being a premium feature and starts being the default way to talk to software. If you are building a voice agent, wire up standby before you ship — your first invoice will thank you.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top