OpenAI has officially opened the gates on GPT-Live-1, exposing its full-duplex speech capabilities to developers via a new API. This isn’t just a text-to-speech wrapper; the model listens and speaks simultaneously, eliminating the awkward pauses that have plagued earlier voice assistants. While the technology already powers ChatGPT’s voice mode, this release lets engineers plug it into their own applications, pairing the speech layer with backend models tuned for specific reasoning depths or cost constraints.
The performance leap is measurable. OpenAI reports that turn-taking latency has dropped to 0.8 seconds, down from 1.4 seconds in its predecessor, GPT-Realtime-2.1. Accuracy metrics show similar gains, with tool-calling precision jumping from 60 percent to 87 percent. In specialized banking support benchmarks, the new model achieved a 32 percent pass rate, more than doubling the previous 12.4 percent score. Yelp is already putting these numbers to work, using the API for phone reservations. CTO Alex Levy notes the change has led to better call handling, suggesting the tech is ready for high-volume customer service roles.
Accessibility comes with a price tag. At $0.05 per minute, the API is not cheap for casual experimentation, signaling OpenAI’s focus on enterprise-grade use cases rather than hobbyist projects. To support global deployment, the update includes twelve new voices covering various accents and dialects, along with built-in ASR transcripts and response text. This allows developers to log interactions for compliance or analysis without building extra infrastructure.
This move effectively commoditizes natural voice interaction for any app willing to pay the premium. By separating the speech interface from the reasoning engine, OpenAI gives developers flexibility but also shifts the burden of optimization onto them. The real test won’t be benchmark scores, but whether businesses can justify the per-minute cost against the value of smoother customer conversations. If Yelp’s early results hold up, we may soon see voice AI moving beyond novelty features into core operational workflows.
