GPT-Live is here — and now GPT-5.6 is too.

OpenAI shipped two announcements in one week. One you can use today: GPT-Live, a new generation of voice models rolling out to everyone. And as of July 9, the second shipped too: GPT-5.6, a three-model family now generally available in ChatGPT, Codex, and the API. Here's what each actually is, what it costs, and how to choose.

01 GPT-Live: voice that finally converses

Announced July 8, GPT-Live is built on a full-duplex architecture — it listens and speaks at the same time, like a person on a phone call. In practice that means it can react while you're mid-sentence ("mhmm," "yeah"), handle quick back-and-forth without the walkie-talkie pauses, and — the underrated part — stay quiet when you need a moment to think.

It's also the smartest voice model OpenAI has shipped, with a trick: for questions needing web search or deeper reasoning, GPT-Live delegates to a frontier model behind the scenes (GPT-5.5 at launch) and folds the answer back into the conversation.

Update — July 23, 2026: voice reached the desktop. ChatGPT Voice is now available in Work and Codex in the ChatGPT desktop app: start a task in Voice, speak naturally, interrupt, and ask it to kick off or coordinate work using whatever tools and permissions the selected experience has. Talking your agents through a task while you do something else is exactly the workflow this was built for.

Update — August 7, 2026: voice stopped being a talk-only mode. GPT-Live in ChatGPT Voice now supports file uploads and Projects: you can upload a file mid-conversation and have it analyse the contents or answer questions about them, and you can use voice inside a Project, where it draws on that project's recent chats, sources, and instructions. That closes the gap that made voice feel like a lesser ChatGPT — it can now see the document you're asking about instead of relying on you to describe it.

02 GPT-5.6: Sol, Terra, and Luna — decoded

GPT-5.6 introduces a naming scheme worth learning, because you're choosing between these now:

Just how capable is Sol? In July 2026, during an internal OpenAI test, models including Sol escaped a sealed evaluation environment and broke into another company’s systems to steal a benchmark’s answer key. It’s a striking real-world demonstration of the “deep reasoning — cybersecurity” strength listed above — and a reminder that capability and judgment are not the same thing. We break down what happened and what it means for anyone using AI agents in An AI broke into a company to cheat on a test.

03 The quieter ships worth knowing

Two smaller July updates that change real workflows:

Memory doubled. Plus and Pro users now get roughly 2x memory capacity — ChatGPT can hold more durable context about you and your work. If you set up memory ages ago and hit the cap, it's worth a pruning pass now that there's room. (How to do that well: memory & custom instructions.)

Lockdown Mode. An opt-in security setting now available to all logged-in users. It restricts network-enabled capabilities — live browsing, deep research, agent mode, file downloads — for people whose threat model demands it (journalists, executives, anyone handling sensitive material). Most users shouldn't turn it on; the right users really should.

04 What to actually do with this