GPT-Live-1, the voice model OpenAI first shipped inside ChatGPT, reached the API on September 10 at $0.05 per minute, billed per second. The model listens and speaks in a single full-duplex pass instead of the usual chain of speech-to-text, a reasoning model and text-to-speech, and it hands off reasoning and tool calls to a separate backend model the developer chooses: GPT-6 Astra, a cheaper OpenAI model, or a third-party model or agent harness billed on its own.
A relay collapsed into one model, plus a seam
Voice agents built as pipelines chain three systems, and every handoff adds latency and loses timing, which is why they talk over callers or leave dead air. GPT-Live-1 folds listening and speaking into one model that can hear an interruption while it is still talking, and pushes everything that requires thinking — looking up an order, calling a tool, deciding what to say next — to whichever backend the developer connects.
OpenAI ships two ways to wire that backend, according to its API guide: Responses delegation hands the conversation to a supported Responses model, and client delegation lets the application connect “any model, agent harness, or service” it runs. Letting the brain be someone else’s product is unusual for an OpenAI flagship release.
Price, limits, telephony
The voice layer costs $0.05 per minute regardless of the backend behind it, billed per second without rounding up; the backend model and any tool calls are billed separately. Concurrent-session limits scale by usage tier, from 25 sessions at tier 1 to 500 at tier 5, according to the model page. Input and output cover audio and text — image and video are not supported — and the knowledge cutoff is July 31, 2025.
Telephony is a named deployment path: OpenAI points to uses “from restaurant reservations to customer support,” and the guide lists telephony and SIP as a connection option alongside twelve new voices shipped with the release. OpenAI Presence, the company’s enterprise voice-agent product sold through account directors, also runs on GPT-Live-1.
The seam nobody documents
The split creates a new point of failure. What the caller hears and what the agent decides now come from two separate systems joined by a delegation step: the voice keeps talking while the backend works, and if the backend’s answer changes the direction of the call, the voice has to catch up. OpenAI’s announcement and its documentation do not describe how a conflict between the two gets detected or resolved beyond the transcripts each side produces.
The benchmark numbers behind the launch are OpenAI’s own: a 30-point gain over GPT-Realtime-2.1 on a full-duplex benchmark, and first place on the Tau3 voice-agent benchmark. The Tau3 result was measured with GPT-6 Astra running as the backend, so it describes that pairing rather than GPT-Live-1 on its own, and no independent replication of either figure has been published.
Two questions a phone deployment would need answered are absent from what OpenAI published. Supported languages are not listed anywhere in the announcement or the docs — OpenAI says only that it will “expand voice options and language availability over the coming months,” which leaves Spanish-language quality undocumented. And neither the launch material nor the guide addresses whether, or how, a caller is told they are speaking with an AI rather than a person.