This was the week where everyone showed up at the same time. Three major model families dropped, Meta changed direction, and Anthropic made a cybersecurity play that nobody expected. Even Microsoft, usually happy to resell OpenAI, showed up with their own models.
The biggest surprise came from Anthropic. Claude Mythos Preview is a general-purpose model that happens to be very good at breaking into things. It found thousands of zero-day vulnerabilities, including a 17-year-old remote code execution bug in FreeBSD. Anthropic is not making Mythos widely available, but 40 organizations including Apple, Microsoft, and CrowdStrike got early access through Project Glasswing.
As someone who works in security, I find it hard to overstate how big this could be if the results hold up.
Meta went the other direction. Muse Spark, their first model under Alexandr Wang’s Superintelligence Labs, is proprietary. No open weights, no Llama tradition. The Meta AI app jumped to #5 on the App Store right after launch, so the market does not seem to mind the closed approach.
Whether this is permanent or just the premium tier of a dual strategy, we should know more at LlamaCon later this month.
On the open side, Google shipped Gemma 4 under Apache 2.0 with sizes from mobile to workstation, handling video, images, and function calling natively. The whole family is already on Hugging Face. Microsoft also showed up with three MAI foundation models for transcription, voice, and images, built by Mustafa Suleyman’s team. Not frontier-scale, but the clearest signal yet that Microsoft wants its own models.
OpenAI spent the week on policy instead of products. Their Industrial Policy for the Intelligence Age proposes robot taxes, public wealth funds, and a four-day workweek. They also confirmed retail investors will get IPO shares when the company eventually goes public. Policy papers from AI labs used to feel like PR. This one reads more like preparation.
T.
