OpenAI has released GPT-Live-1 through its API, a speech model built so that an application can listen and speak at the same time. On OpenAI's own benchmarks it scores 80.1% on full-duplex interaction against 45.4% for GPT-Realtime-2.1, cuts turn-handover latency from 1.4 seconds to 0.8, and lifts tool-calling accuracy from 60% to 87%. It ships 12 new voices spanning accents, dialects and languages, and returns ASR transcripts alongside the response text rather than making developers bolt on their own. The numbers are large, and one of them is much smaller than the rest.
Source: the-decoder.com
Full duplex is the headline claim and the one that changes what a voice app feels like. A half-duplex system is a walkie-talkie: it talks, you wait, you talk, it waits. Every interruption is a failure state. Pair the 80.1% duplex score with 0.8 seconds of handover latency and you get the two properties that, to my mind, separate a conversation from a phone tree. OpenAI has put the details in its API documentation. The jump from 60% to 87% on tool calls matters for the same reason from the other direction: a voice agent that cannot reliably invoke the right function is a narrator, not an assistant.
Then there is the bank benchmark. On OpenAI's test of voice support for banks, GPT-Live-1 returns 32% successful responses, against 12.4% for the previous model.
That is the number worth sitting with, and it is the only one in the set that looks like an end-to-end job rather than a component measurement. Interpreted charitably, it is a 2.6x improvement, and on a base of 12.4% that is a real leap. Interpreted the way a bank's operations lead will interpret it, GPT-Live-1 fails roughly two out of every three support interactions it is asked to handle. The gap between 80.1% and 32% is the gap between a model that converses well and a system that finishes work, and OpenAI published both, which is more candor than this category usually gets.
My read is that the release is honest about the state of the technology and the marketing around it will not be. Full duplex demos beautifully. A model that talks over you less, responds in under a second and does twelve accents will sell itself in a thirty-second clip, and the clip will not be running a banking workflow. Also worth flagging: these are internal benchmarks. There is no independent scoring here, no definition given of what counts as a successful response in the banking test, and nothing on what any of this costs, which for a model designed to sit on an open microphone is the number that decides whether an always-on listening app is viable at all.
The strange consequence of shipping this now is that the better the conversation gets, the more expensive the failures become. A bot that sounds like a phone tree sets expectations a phone tree can meet. One that interrupts naturally, answers in 0.8 seconds and speaks your dialect invites a customer to explain a complicated problem in full, and then resolves it 32% of the time.