OpenAI launches GPT-Live, a voice model built for more natural conversation than its predecessor

OpenAI announced immediate availability of GPT-Live-1 in its API after a test series that included 80 real restaurant reservation calls. According to the released figures, the task completion rate more than doubled versus the previous voice model, reaching 92% perfect understanding. Those metrics come directly from OpenAI, not from an independent third-party benchmark.

The core improvement lies in interruption handling. The model no longer halts when the user emits a quiet acknowledgment such as "hmm," and it can switch topics mid-sentence without perceptible lag. A "listen while speaking" capability lets the conversation flow closer to ordinary human dialogue, dropping the forced "turn-taking" that characterized earlier models.

The model is already live on the Genspark platform across three products: Call for Me, Realtime Voice Agent, and a new addition called GenTeam. GenTeam lets users call automated agents as if they were teammates — delegating tasks, checking status, and retrieving results through an interface that simulates an internal phone call rather than a text-based chat.

GPT-Live-1 is now accessible through the OpenAI API, letting developers embed continuous conversation capabilities into their own applications. The documentation notes support for models and tooling (harness) of the developer's choice, without mandating a closed toolchain. No usage pricing or specific rate limits for the new model have been published yet.

What the announcement leaves out: no comparison to competitors such as OpenAI's own previous Realtime API, or to voice models from Anthropic and Google. No indication whether testing covered noisy environments, non-standard accents, or conversations longer than a single reservation call. Until independent benchmarks appear, the numbers remain manufacturer's data.