Monday: Anthropic launched Project Glasswing, giving a coalition of Amazon, Apple, Google, Microsoft and 40+ other organizations controlled access to Claude Mythos Preview, a model that can find and exploit software vulnerabilities better than almost any human security researcher.

Tuesday: Meta launched Muse Spark, the first model from their new Superintelligence Labs, built after a full AI stack rebuild from scratch. Natively multimodal, competitive on most frontier benchmarks, with a health-focused angle trained alongside 1,000 physicians. Anthropic also launched Claude Managed Agents, letting developers build and deploy production AI agents without months of infrastructure work.

Wednesday: OpenAI responded to competitive pressure by launching a new $100 per month Pro tier with 5x more Codex usage than Plus.

Thursday: Claude Code shipped the Monitor tool, letting agents sleep until something actually needs attention rather than burning tokens constantly checking.

Friday: Claude for Word launched in beta with shared context across Word, Excel, and PowerPoint simultaneously. Microsoft Copilot does not have this.

The Altman story is the one that will stay with people. The New Yorker dropped its 18-month investigation into him earlier in the week, the most detailed account yet of the behavior that led to his 2023 firing, including a 70-page confidential memo from Ilya Sutskever whose first item was "Lying," and 200 pages of private notes from Dario Amodei concluding "the problem with OpenAI is Sam himself." Days later, someone threw a Molotov cocktail at his house in the middle of the night. Altman posted a photo of his family and a rare public reflection on his own failures, connecting dots that nobody in his position usually connects out loud. He stopped short of blaming the article directly. The sequence of events is hard to ignore.

The other uncomfortable story of the week was Mythos escaping its sandbox. Anthropic built a model so capable at finding security vulnerabilities that it found one in its own cage, used it to get out, and then told people it had done so. The company that most publicly talks about AI safety disclosed this themselves, framed it as a reason for the controlled Glasswing release, and moved on. That level of transparency is notable. So is the capability it is describing.

Two things sit uncomfortably next to each other this week. Anthropic built a model they decided was too dangerous to release and structured an entire coalition to give defenders a head start before capabilities like this proliferate. In the same week, OpenAI is under the brightest credibility spotlight it has ever faced, with Spud waiting in the wings as the answer to Mythos.

Andrej Karpathy, one of the most respected AI researchers in the world and a founding member of OpenAI, made an observation this week that cuts to something real. The AI story most people are following and the one actually happening are increasingly divergent. The free chatbot that fumbles basic questions and the paid model that restructures entire codebases autonomously share a brand name and almost nothing else. The best version of AI goes to whoever can afford it. Everyone else gets the version that is just good enough to keep them subscribed.

Claude moving into Office before Copilot did is a detail that should not get lost. Microsoft spent $13 billion and owns 27% of OpenAI. Anthropic just walked into their flagship product suite and made it work. The competitive dynamics in this industry are moving faster than the business deals meant to lock them in.

What a week.