GPT-Live 1 Comes to the API: Can It Take Customer Service Calls, Whose Voice Is It, Will It Record?

On September 10, 2026, OpenAI’s API developer documentation logged a new changelog entry: the full-duplex voice model GPT-Live 1 is now generally available in the API, under the model ID gpt-live-1 and the endpoint v1/live/sessions. The official documentation states that voice sessions cost $0.05 per minute, billed per second, with backend model and tool usage charged separately. What this means is that a voice model that can listen and speak at the same time is a component developers can buy and build into their own phone systems, apps, or products.

This article was checked on September 18, 2026 against OpenAI’s developer documentation: the API changelog, the GPT-Live 1 model page, the telephony and SIP integration guide, and the session management guide, reflecting the versions available on the day of checking; these documentation pages carry no version numbers and their content may change later. This site did not actually integrate with or test this API; the capability descriptions in this article are OpenAI’s own statements. This article covers only this update on the API developer side, not the plans or settings on the ChatGPT consumer side.

The September 10 Changelog Entry: Generally Available, the Model ID, and Two Delegation Modes

The official model page defines GPT-Live 1 as a full-duplex voice model: it can listen and speak at the same time, delegating reasoning and tool use to a backend agent — meaning that when something needs to be looked up or computed, that work is handed off in the background while the conversation continues. The changelog describes the September 10 status as “now generally available in the API”; that phrase describes an availability status, and the four documents this article checked do not say this is the first time it has appeared. The telephony integration guide, in fact, notes that existing integrations may still receive a deprecated incoming-call event name, which suggests integrations were already running before this date.

The model page lists the model ID as gpt-live-1, with the corresponding endpoint v1/live/sessions; the official documentation describes it as “our premier model for natural, expressive voice conversations with smooth interruption handling” — this is OpenAI’s own characterization. Both input and output modalities are audio and text; image and video are not supported. The model page lists a knowledge cutoff of July 31, 2025, which is the point in time for the model’s own training data.

When developers want GPT-Live 1 to handle questions that require looking things up or computation, the documentation describes two delegation modes: Responses delegation, paired with an OpenAI model, and client delegation, which connects to a backend the developer runs themselves; if delegation is omitted or set to null, it defaults to client mode. The documentation states that the delegation mode cannot be changed once a session has started — changing it requires starting a new session.

The Model’s API Specifications: Only One Supported Endpoint, No Free Tier

GPT-Live 1 has only one endpoint marked as supported in the API — Live (v1/live/sessions); the official endpoint list marks existing endpoints such as Chat Completions, Responses, and Realtime, along with the older Completions (legacy), as not supported. The telephony integration guide also notes that each API has its own authentication, session creation, and event contract, and they cannot be substituted for one another.

Usage is also measured differently: the documentation states that rate limits are measured in concurrent sessions, and lists Free as an unsupported usage tier. Depending on the account’s usage tier, Tier 1 through Tier 5 can open 25, 50, 200, 300, and 500 concurrent sessions respectively.

When a Call Comes In: How GPT-Live 1 Takes Over

To connect GPT-Live 1 to a phone system, the official documentation lists two paths: Direct SIP, where the telephony provider exchanges call audio with OpenAI while the developer’s application handles webhooks, session configuration, call decisions, and business logic; and a server audio bridge, where the developer’s application relays audio from the telephony provider or a meeting room to GPT-Live over WebSocket, managing both connections, event translation, playback, and the call lifecycle itself. The documentation mentions that Twilio, Telnyx, LiveKit, and Daily/Pipecat each have dedicated integration guides.

Once a call comes in: the developer’s application applies its own authorization and routing rules, using a separate API action to accept or reject the call — accepting requires sending settings such as the model, voice, and delegation mode, while rejecting requires attaching a SIP status code between 300 and 699 (486, for busy, for example); the first accept or reject decision wins, and any decision sent afterward is rejected. Two further actions are available during a call — transferring it elsewhere or hanging up directly — both of which return 200 OK with no content on success.

At the end of this section, the documentation states plainly that this flow handles inbound calls, and does not support creating outbound SIP calls through POST /v1/live/sessions; for outbound calling, the documentation directs developers to use a partner integration instead, describing outbound calling as something the provider owns. The documentation also warns that SIP headers attached to an incoming call should be treated as untrusted caller metadata, not as a basis for authorization.

Whose Voice Is It: 12 Built-in Options, but These Four Documents Don’t List Supported Languages

GPT-Live 1’s voice isn’t limited to just one option. Besides the default, marin, the official documentation lists 12 additional named voices to choose from — for example, vesper, whose language field is marked English and regional-influence field British, and gleam, marked English and North American; as well as bossa and tempo, both marked Portuguese with a Brazilian regional influence. Developers select one and put it in the session settings when creating a session; once a conversation has started, it cannot be changed mid-call — changing it requires starting a new session.

What readers in Taiwan should note: the documentation uses “regional influence” to describe these voices, while specifically stating that this refers only to speaking style and is not a guarantee of accent fidelity. More importantly, the four official documents this article checked do not list which spoken languages this model supports — the language field for the 12 named voices shows only English and Portuguese, and the default voice, marin, has no language specified either; when the documentation discusses greetings, it tells developers to use the language specified by the application until the caller speaks, and to test with the languages their own application supports. In other words, if GPT-Live 1 really is answering a customer service call, how well it handles Chinese is a question these four documents do not answer — it should not be assumed that it can.

The official documentation also mentions that, once approved, a custom voice can be built from your own recordings; the details are in a separate guide that this article does not go into.

Is It Recorded, How Long Does It Remember, and What Does a Few Minutes on the Phone Cost

Whether a call is recorded is up to the developer — it is not on by default. The official documentation states that a session’s recording setting defaults to off; the developer must turn it on, and the project must first be approved, before a completed recording can be kept for download or “forking” into a new session, and downloading or forking also requires a data policy that permits persistence. Saved recordings expire after 30 days; if a project has Zero Data Retention enabled, the recording setting is always treated as off, making downloads and forks unavailable. A downloaded recording is a stereo WAV file, with the input and output audio in the left and right channels respectively.

There’s also a cap on memory for long calls. The documentation states that a session’s default context window holds 128,000 tokens, which includes the instructions given by the developer, the conversation text, and audio tokens that do not appear in the transcript; when usage exceeds 90% of that window, the system starts a replacement voice engine within the same session, and the new engine receives the original instructions plus up to 8,192 tokens of conversation history (recent messages, and, when available, a summary of older messages). That means that in a very long call, details discussed earlier may be summarized or omitted.

The pricing formula itself is simple: voice sessions cost $0.05 per minute, billed by actual seconds, and the documentation notes that this is not rounded up to the next whole minute; the backend model and tools are billed separately at their own rates, and these four documents do not combine the two into a single worked example. What follows is this article’s own conversion at the rate checked on September 18, 2026, not an official figure: for three full minutes of voice alone, that works out to $0.15; the actual total also depends on how much work the backend did. When a session ends, the official documentation lists reasons including the application closing it, the session reaching its duration limit, a safety filter stopping it, the remote connection ending normally, and the connection being interrupted unexpectedly — but the documentation does not state what the duration limit actually is in minutes.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

Gemini 3.8 Live and Extended Thinking: New Voice Models, Can Your Account Use Them?

• AI tools
• Apps & software
• AI news