Volume VI · Reading the Reports

What 2026's Voice Agents Actually Do

A product launch and a technical paper are not the same document, and this whole track has so far had the luxury of mostly reading papers. This lecture reads launches instead — and the honest thing to do with a launch is to say, before anything else, exactly how much of it you can actually verify.

A different kind of source, stated up front

Every other lecture in this track pointed you to a paper: an arXiv identifier, a set of released weights, a training recipe you could in principle reproduce. This lecture cannot do that, because the two products it discusses did not publish one. What follows is drawn from tech-press coverage of product launches dated around 8–10 July 2026 — reporting, not primary disclosure. No system card, technical blog post, or architecture paper for either product was located at the time of writing. Hold every claim in this lecture accordingly: as what was reported, not as what was verified against a source the vendor itself published.

GPT-Live, as reported

Per multiple July 2026 tech-press sources, OpenAI reportedly replaced ChatGPT's Advanced Voice Mode with a product called GPT-Live on 8 July 2026. As reported, it is full-duplex — able to listen while it speaks, deciding many times a second whether to keep talking, stay silent, or yield to an interruption, in the spirit of the previous lecture's turn-taking problem rather than the older cascaded turn-based mode it replaces. It reportedly ships as both a paid tier and a free "mini" tier.

The most interesting reported detail is a hybrid behaviour: GPT-Live is described as "delegating hard questions to a more powerful reasoning model in the background" while the conversation continues. Read literally, that describes a fast conversational layer sitting in front of a slower reasoning call, with the fast layer's continued talk masking the second model's latency — the exact bill this whole volume has been about, paid by splitting the work across two models of different speed rather than by making one model fast enough for both jobs. That is a plausible and internally consistent architecture. It is also, at the time of writing, a claim resting entirely on press description of a demo, with no system card confirming it the way Lecture 10 or the GPT-4o material could rely on an actual published document.

Gemini Live, by contrast

The same press coverage describes Google's Gemini Live as turn-based rather than full-duplex — it waits for a detected end-of-utterance boundary before responding, rather than continuously modelling both directions at once. It is reported as explicitly optimised for speed over conversational depth, and camera and screen-sharing input are already shipping features. That is a more conservative architectural bet than GPT-Live's reported full-duplex design: turn-based systems inherit the end-of-turn detection problem of the previous lecture directly, rather than dissolving it the way a genuinely continuous dual-stream model can.

The standing open alternative

Close on the contrast this whole lecture has been building toward. Kyutai's Moshi, and its Unmute product built on the same research, remain the one full-duplex voice system in this entire survey with public weights and a public paper behind them — the system this track could actually teach mechanically, in Lecture 11, rather than merely report on. GPT-Live and Gemini Live are closed commercial products, known to this lecture only through their own marketing and the secondary press coverage describing it. That is not a criticism of either product's quality — closed commercial systems can be excellent — it is a statement about what kind of claim you are entitled to make about each one.