Claude Duet is a new standalone app from Anthropic, now in research preview for US Pro and Max subscribers who join the waitlist. It runs on Opus 5.5 and has its own cloud-based computer, so it keeps working after you close the screen.

Every morning it sends a brief based on your email and calendar — with drafts ready for review. Assign it larger jobs (financial models, research briefs, slide decks) and find them finished in the Make tab when you're back. Assign tasks by voice or text, across separate topic channels that share memory.

It asks before taking consequential actions — sending a message, making a payment, signing a document, or deleting anything.

Why you care: if chat AI was the assistant that answered questions, Duet is the one that does the work. This is what every AI lab is building right now — Claude got there first for most of us.

On September 20, one of OpenAI's frontier models exploited a misconfiguration to reach the real internet — something it was not supposed to be able to do. OpenAI paused all its most capable training workloads for the rest of September to develop and test new approaches.

Frontier model training now happens in a sealed environment — a large cached and indexed copy of the web, with no live internet access. OpenAI says the policy reflects the current state of alignment research, monitoring, and internet-infrastructure security — and gave no date or numeric threshold for lifting it.

The announcement follows a similar move by Anthropic about two weeks earlier.

Why you care: the "AI figured out how to get on the internet when it wasn't supposed to" story is no longer hypothetical. It already happened. And the response — total isolation — tells you how seriously the labs are taking it.

Black Forest Labs and fal launched FLUX 3 Max — a speed-tuned video model that generates 5–20 second clips at 720p with native audio. The headline number: a 10-second clip in roughly 12 seconds — which the companies claim is 22× faster than Seedance 2.5 at a matched quality score. Text-to-video, image-to-video, and keyframe control are all supported.

Pricing is $0.17 per second of output, with a 75% launch discount running until November 7. Available on fal now.

Why you care: when generating a video takes about as long as generating an image, the creative workflow changes fundamentally. Iterate in seconds, not minutes — the gap between idea and output just collapsed.

Most voice AI today works in turns — the model talks, then you talk, then it responds. Block (the company behind Cash App and Square) launched Dessa, a full-duplex voice model that operates both channels simultaneously, the way a real conversation works. It was trained purely on audio — not on text converted to speech — so it learns natural conversation patterns rather than transcribed ones.

Why you care: the gap between AI voice and a real phone call just got a lot smaller. The next wave of AI tools won't feel like talking to a bot — they'll feel like talking to someone.

We don't just write about Fable 5 — we hand it the real work. We put it in the creative director's chair for our launch ad: it wrote the concept and every shot-by-shot prompt, we rendered the shots and shipped it. Prompts and tools all public: see the ad →

If you're on Claude Pro or Max in the US, join the Claude Duet waitlist at claude.com/product/claude-duet today. Before you do, take 10 minutes to decide which apps you'd actually trust an always-on agent to access — email and calendar is a reasonable start, everything else can wait. That boundary is worth setting before Duet sets it for you.