OpenAI split a voice model’s brain. Then one team deleted 23,000 lines of code.
The New Stack Amanda Caswell ● Covered by 2 sources
OpenAI put its live voice model in the API. It can hand hard work to another model instead of making callers wait in silence.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Building a voice agent has always been messier than the demos make it look. Usually, speech turns into text, text goes to a model, the answer gets turned back into audio, and the whole thing can start sounding like a robot with a bad delay problem. OpenAI is trying to flatten that pipeline with GPT-Live-1, which it launched in its API on Wednesday.
The pitch is simple: let one model stay in charge of the conversation while a heavier model does the thinking behind the curtain. GPT-Live-1 is the front line. It’s natively full-duplex, so it can keep talking through interruptions and mid-sentence cut-ins instead of waiting for each turn to be perfectly tidy. And when a request needs more work, it can delegate the task elsewhere rather than grinding the exchange to a halt.
That backend model does not have to be OpenAI’s own. OpenAI says GPT-Live-1 can hand off to GPT-6 Astra, a smaller model like Luna, or even a model from another provider. The company exposes that handoff through events: the voice session creates a delegation_id, passes context to the backend system, then receives the result through session.commentary.append and folds it back into the live conversation. Developers still get visibility and control over turns. They just stop wiring together every piece of the stack themselves.
OpenAI is also leaning on results to make the case. It says GPT-Live-1 beats GPT-Realtime-2.1 by 30 percentage points on Full Duplex Bench. Paired with GPT-6 Astra at medium reasoning, it also takes the top spot on the 𝜏³-benchmark. Early users seem to care less about the benchmarks than the cleanup. EliseAI said one team deleted 23,000 lines of code after switching, an 80% cut, while Speak found the model was nearly 80% less likely to jump in when a learner paused to think. Yelp says it is already seeing more calls handled successfully, with callers using fuller, more natural sentences.
The price is the usual platform tax. GPT-Live-1 costs $0.05 per minute, about $3 an hour, before whatever model sits behind it gets billed too. That is the tradeoff OpenAI wants developers to accept: less plumbing, more dependence. If the voice layer is good enough, plenty of teams will take the deal and stop pretending they enjoy maintaining 23,000 lines of glue code.
My take — AI-written commentary, not fact-checked reporting
This is the kind of move that makes closed platforms look irresistible right up until the lock-in bill shows up. OpenAI is betting developers will trade control for less clunky conversations, and that bet usually works — especially when the glue code is ugly enough to qualify as a cry for help. The real trick is whether teams notice they’ve handed over the front door before they’ve finished celebrating the cleaner codebase.
Read more about this at: The New Stack
