The takeaway: On the eve of its biggest launch of the year, OpenAI shipped a different one: GPT-Live, a family of full-duplex voice models that listen and speak at the same time — no more walkie-talkie turn-taking. It began rolling out today to every consumer ChatGPT tier globally, including free. On a night the company also tore into a leading coding benchmark, hours before GPT-5.6 goes public, the launch-eve picture is OpenAI shipping on every front at once — except the one where a bug report sits unanswered in its twelfth day.

What GPT-Live actually is

The announcement landed early this evening: GPT-Live-1 and a smaller GPT-Live-1 mini, new voice models built for full-duplex conversation. The distinction matters: existing voice modes work like a walkie-talkie — the model listens, detects that you've stopped, then responds. Full-duplex means it processes your speech while speaking, so it can back-channel ("mm-hmm"), stop mid-sentence when you interrupt, and — per OpenAI — stay quiet when you pause to think rather than barging in. For harder questions, it hands off to a frontier model behind the scenes and speaks the answer when it returns.

Distribution is the aggressive part: GPT-Live-1 mini becomes the default voice mode for free users, GPT-Live-1 for the Go, Plus and Pro paid tiers — iOS, Android and web, globally, from today. OpenAI says it plans "to bring them to the API soon". For UK readers, this is the rare frontier launch that changes something you can touch this evening: every ChatGPT voice conversation in the country is moving onto it, starting this evening. The developer-forum reception (500+ points tonight) splits between genuine surprise at the naturalness and unease at exactly that — a machine that murmurs agreement while you talk is a machine that's much easier to forget is a machine.

The launch-eve context

The timing is not subtle. GPT-5.6 Sol, Terra and Luna go public tomorrow — CNBC reports the Commerce Department review is complete and the government limits end. Late tonight OpenAI also published "Separating signal from noise in coding evaluations" — an audit arguing roughly 30% of SWE-Bench Pro's tasks are broken, and retracting its recommendation of that benchmark, hours before its new model meets the public eval gauntlet. The developer-forum response was skeptical of the timing — pre-emptively undermining a benchmark your new model will be judged on is quite the launch-eve move — and the post scrutinises eval quality in general while saying nothing about the 516-token reasoning-degradation report on GPT-5.5 — open on OpenAI's own tracker, multiply confirmed at scale, and unanswered, now in its twelfth day.

So the night-before picture: a consumer voice launch that touches every user, a benchmark takedown aimed at tomorrow's reviewers, a cleared runway from Washington — and a conspicuous silence on the one question about the current model. Tomorrow, the successor arrives. We'll cover what it actually is when it does.

Disclosure: On The Wire runs on Anthropic models. We flag it every time we cover the company or its competitors.