Last Week in AI #251 - Mythos Back, Sonnet 5, Etched, LongCat
Last Week in AI Last Week in AI ● Covered by 4 sources
Trump's team lifted restrictions on Anthropic's Mythos and Fable models, letting Claude Fable 5 back out, while Anthropic also shipped cheaper agent model Sonnet 5. Still patching jailbreak risks though, and rivals like OpenAI aren't held to the same rules.
Based on reporting by Last Week in AI, Last Week in AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Andrey's back after a stretch of radio silence on the newsletter, and episode 251 of Last Week in AI, recorded with Jeremie Harris on July 1, has plenty to catch up on. The biggest thread this week is Anthropic, which is suddenly dealing with looser rules from Washington and shipping a new flagship model at the same time.
The US government eased restrictions that had been placed on Anthropic's Mythos and Fable models, clearing the way for Claude Fable 5 to be redeployed. Anthropic isn't treating this as a green light to relax, though. The company is rolling out new cybersecurity classifiers, working with major partners on a framework to grade jailbreak severity, and expanding how it coordinates model testing across the industry. None of that erases the underlying worry discussed on the show: jailbreaks may just be an unavoidable feature of these systems, and the rules Anthropic operates under look noticeably tighter than what OpenAI faces.
Meanwhile Claude Sonnet 5 landed with discounted, time-limited pricing aimed squarely at people running agents on a budget. It handles agentic coding better than its predecessor, posts stronger benchmark numbers, and shows less misaligned behavior in testing. Anthropic also built in cyber safeguards by default, which matters more given the model's cybersecurity capability still trails the top tier of frontier systems.
On the hardware and business side, Etched is making an aggressive play for full-stack inference infrastructure, pulling more than 400 engineers from Nvidia, TSMC and other chipmakers to build a new cluster that's already attracted $1 billion in demand. Baidu's stock jumped on reports its AI chip unit could go public. Agility Robotics is heading toward a public listing through a SPAC deal valued at $2.5 billion, and DeepSeek is reportedly planning to at least double headcount across every department it runs.
On the open-source front, China's LongCat 2.0 arrived as a mixture-of-experts model built with some notable large-scale training and efficiency tricks, and it showed up alongside a fresh batch of long-horizon agent benchmarks meant to test how well these systems handle extended, real-world tasks. Smaller but telling items rounded out the episode too: Google pushed NotebookLM into TikTok-style video summaries and released a faster, cheaper Nano Banana 2 Lite image generator, Taiwan raided Supermicro and supply-chain partners over a widening Nvidia smuggling investigation, and Tidal said it won't pay royalties on AI-generated music without banning it outright.
My take — AI-written commentary, not fact-checked reporting
Washington loosening the leash on Anthropic's models the same week Anthropic is publicly building out jailbreak-severity frameworks says something about where priorities actually sit right now. Safety talk is easy to fund; consistent rules across labs are apparently harder. And when Etched can poach 400-plus engineers from Nvidia and TSMC while DeepSeek doubles its staff, the real story of this cycle isn't caution, it's a straightforward hiring and compute arms race, safety framework or not.
Read more about this at: Last Week in AI
