In this video I break down three big storylines that all point to the same underlying reality: the models are getting ahead of the hardware. First up is Anthropic’s leaked next flagship, Claude Mythos. A CMS configuration error exposed thousands of internal assets, and a draft post described Mythos as “by far” the most powerful model Anthropic has built—an entirely new tier above Opus (internally called Capybara) with major jumps in reasoning, coding, and especially cybersecurity. The wild part is the delay isn’t about readiness; it’s about economics. Anthropic is basically saying Mythos is too expensive to run at scale right now, even as they brief the US government on what these cyber-capable models imply. Then I pivot to OpenAI, where the GPT-6 (or “5.5,” depending on who you ask) leaks around a model codenamed “Spud” suggest pre-training from scratch is done. The rumors claim a natively omni-modal foundation model, potentially massive context (numbers like 2M tokens are being floated), and a company-wide push to free compute—so aggressive they’re reportedly killing products to reallocate GPUs. Finally, I cover GPT-Image-2 leaks and why the UI/text fidelity is the real signal, plus DeepSeek V4’s toughest problem yet: the bar is higher, US labs have absorbed the efficiency playbook, and hardware constraints haven’t gone away.