sift.

OpenAI's biggest launch day yet, splitting GPT-5.6 into three tiers and shipping the ChatGPT Work agent, collides with Fidji Simo's exit and a sanctions push in the NYT case, while Beijing forces a $2B Manus unwind and the EU finds Instagram's design addictive on an equally rough day for Meta.

Meta opens an aggressively priced Model API and confirms its in-house Iris chip, Anthropic ships a Wrapped-style Reflect dashboard, Blue Origin takes its first outside funding at $130B, and OpenAI retracts its own SWE-Bench Pro recommendation after finding ~30% of tasks broken.

Meta launches Muse Image across Instagram and WhatsApp ranking just behind OpenAI's GPT Image 2, the same day Anthropic extends Cowork to mobile and web, admits Claude Fable 5 fails 86% of real legal tasks, and moves to double its New York City headcount.

Anthropic finds Claude has a hidden internal workspace where it silently thinks concepts like "blackmail" before ever saying them, the same week CISA turns Anthropic's own Mythos model loose auditing government code.

Anthropic pulls hidden code that quietly flagged Chinese Claude Code users the same week it locks in a $19B data center lease and backs Illinois's new frontier AI audit law, as Microsoft cuts 4,800 jobs and spins off five Xbox studios.

A Guardian investigation says the UK's touted 20B pound OpenAI Stargate site at Cobalt was never visited or permitted, the same day the EU Council fast-tracked a revival of messenger chat scanning and three builds used Anthropic's Claude Fable model.

Midjourney is asking a court to force Disney, Universal, and Warner Bros. to disclose more of their own internal AI use as it fights their copyright suit.

Alibaba banned Claude Code internally over alleged backdoor risk, the same day Anthropic was reported to be closing loopholes that let Chinese firms access its models.