OpenAI ships GPT-Live-1 speech API at $0.05 a minute, doubling interactivity score

OpenAI has released GPT-Live-1, a speech model that listens and talks at the same time, as a public developer API. The company says the model scores 80.1 percent on interactivity evaluations, nearly double the 45.4 percent posted by the system it replaces. Pricing is set at $0.05 per minute of use, a rate that puts sustained voice sessions well above the cost of most text-based calls.

The jump in interactivity matters because full-duplex behavior is the hard part of voice AI. Older speech systems typically wait for a user to finish before responding, which produces the familiar pause-and-reply rhythm of a walkie-talkie. GPT-Live-1 handles input and output simultaneously, so it can react to a speaker mid-sentence, backchannel, or be interrupted without losing the thread. That capability is what separates a natural phone call from a stilted menu tree.

OpenAI has been moving toward real-time voice for more than a year, and the API release pushes the technology out of its own consumer apps and into third-party products. Developers building customer support lines, language tutors, and in-car assistants have so far relied on stitching together separate speech-recognition and text-to-speech services. A single model that covers both directions removes a layer of orchestration, plus the latency that came with it.

The price is the catch. At $0.05 per minute, an hour of conversation costs about $3, and a service running thousands of concurrent calls can burn through budget quickly. Competing speech APIs from Google, Amazon, and a cluster of startups generally price below that for standard recognition and synthesis, though few offer true simultaneous two-way audio. Teams will have to weigh the interactivity gain against per-minute spend, especially for always-on applications.

OpenAI has not published the benchmark methodology behind the 80.1 percent figure, and independent evaluations of full-duplex systems remain scarce. The practical test will come from developers: whether apps built on GPT-Live-1 feel noticeably more human than the turn-taking assistants already in the market, and whether the per-minute meter keeps them running long enough to find out.