GPT-Live-1 lands in the API as its own Live endpoint, not a Realtime upgrade

OpenAI put GPT-Live-1 in the API on September 10, 2026. The model ID is gpt-live-1. It listens and speaks at the same time, then hands deeper work to a separate backend so the conversation can keep going while tools run.

That is a new surface. The model page lists support only on v1/live/sessions. Chat Completions, Responses, and the Realtime API are marked not supported for this ID. If you already built on gpt-realtime-2.1, this is a migration, not a drop-in model swap.

OpenAI Developers described it as ChatGPT’s back-and-forth for apps, with voice agents that “listen while they speak and work with the models and harness you choose.” GPT-Live first showed up in ChatGPT Voice. The API release is the developer product: you pick the backend, the tools, and the voice prompt.

Voice in front, agent in back

OpenAI’s voice-agent guide now splits three architectures. GPT-Live is the full-duplex front end with a separate backend. Realtime still uses one model for speech, reasoning, and tool selection. A chained STT-LLM-TTS pipeline stays available if you need to inspect every text stage.

Live sessions make that split concrete. The live model handles turn-taking, interruptions, and when to ask for help. Delegation then goes one of two ways. Responses delegation lets GPT-Live call a configured Responses model and pass conversation context. Client delegation lets your app run any model, agent harness, or service and send a verified result back. You choose the mode when you create the session. Switching modes later fails; start a new session.

The launch post names GPT-6 Astra as a reasoning backend and says third-party models are allowed. The delegation docs start from gpt-5.6-terra, or Luna for cheaper volume work. Your application still owns permissions, confirmations, and durable task state. Interrupting speech does not automatically cancel backend work.

GPT-Live is not a turn-based model, but it still exposes turn detection, ASR transcripts, and response text. WebRTC is the browser path. WebSockets cover server audio. SIP is documented for phone agents.

What $0.05 a minute actually buys

Pricing lists GPT-Live-1 at $0.05 per minute, billed per second, with no round-up to a whole minute. That is only the voice session. Backend tokens and tools bill at the chosen model’s normal rates. A five-minute call that never delegates is $0.25. A five-minute call that keeps Astra thinking in the background is $0.25 plus Astra.

Rate limits are concurrent sessions, not tokens: 25 on Tier 1, up to 500 on Tier 5. The Free usage tier is unsupported. Knowledge cutoff on the model card is July 31, 2025. Input and output are audio and text. Image and video are unsupported. Streaming and function calling are listed; structured outputs and fine-tuning are not.

ChatGPT also shipped GPT-Live-1 mini for Free users. The API model page lists only gpt-live-1.

OpenAI’s scorecard, not an independent rerun

OpenAI says GPT-Live-1 improves Full Duplex Bench by 30 percentage points over GPT-Realtime-2.1, with gains in turn-taking latency and interactive behavior. Paired with GPT-6 Astra at medium reasoning effort, it ranks first on Tau3 in that same post. Those are vendor-reported results. The launch page also describes spoken customer-service and banking tasks, and tool-use tests with Terra at low effort, without printing Pass@1 scores for those rows.

Customer quotes on the page are similarly scoped. Speak said early evaluations cut interruptions during thinking pauses by almost 80% versus previous turn-based systems. That is Speak’s comparison, not an arena rerun.

Custom voices go through sales. Presence can sit on top of GPT-Live for enterprise agents; it is not the self-serve API.

The practical test is local. Put gpt-live-1 on the Live endpoint, keep your existing agent behind client or Responses delegation, and measure barge-in, backend cost, and whether an interrupt leaves a tool call running. Realtime remains the one-model path. Live is the split: a $0.05-per-minute mouth, and whatever brain you can afford behind it.