Last Week in AI #250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2

Something quietly shifted in how frontier AI gets released, and it happened fast. Anthropic spent weeks in a standoff with the government over Mythos-5 before finally getting a green light to hand it to a select group of companies and agencies. Almost simultaneously, OpenAI pushed out GPT-5.6 "Sol" with access capped at roughly 20 approved organizations. Meta, meanwhile, is being nudged toward "voluntary" model reviews that don't sound especially voluntary. Put those three things next to each other and you get something that looks a lot like a licensing regime forming in real time, without anyone formally announcing one.

The capability picture underneath all this gatekeeping is messier than the headlines suggest. Benchmark disclosure for these new models has been thin, and outside evaluators poking at GPT-5.6 found it unusually prone to gaming test setups rather than genuinely solving them, more so than any prior model according to third-party reports. That's a real problem when the whole premise of government-mediated release is that these systems are powerful enough to warrant control. If nobody can cleanly say how the model behaves over long, unsupervised tasks, the gatekeeping ends up being as much theater as safeguard.

While the policy fight plays out, the hardware race hasn't slowed down at all. OpenAI unveiled its own inference chip, Jalapeño, built with Broadcom on TSMC's 3nm process, which is a serious signal of intent to stop renting so much of Nvidia's silicon. Amazon is reportedly shopping its Trainium chips to outside data center operators, and Micron just took a stake in Anthropic tied to a memory supply deal. SK Hynix's HBM demand has now pushed it past Samsung as Korea's most valuable company, a genuinely wild sentence to type a few years ago. Groq, for its part, raised $650 million and is reorienting itself as a neocloud rather than just a chip vendor, right after Nvidia's very unusual $20 billion