This week’s big AI news came in two flavors: new models, and new ways those models can go wrong. Anthropic shipped Claude Opus 5, pitched as nearly as capable as its top model but at roughly half the price. It is aimed squarely at teams who found the top-tier model great but too expensive to run everywhere. In the same week, its own researchers showed just how far that same capability cuts both ways.
Anthropic said its researchers used Claude Mythos Preview to find new attack methods against cryptographic algorithms, real mathematical weaknesses, not a marketing demo. A day later, Simon Willison flagged something less comforting: a self-replicating prompt injection worm that spreads through Word documents via Microsoft Copilot, no attacker needed after the first infection. Same underlying skill, pointed in opposite directions, in the same seven days.
Open models kept moving too, quieter but just as real. Moonshot AI published the technical report for Kimi K3, a 2.8 trillion parameter open-weight model claiming frontier-level results on coding and agentic tasks. Hugging Face released LFM2.5-Encoders, small encoder models built to stay fast as documents get longer, running on CPU alone. Not exciting as a headline, but exactly the kind of release that makes real products cheaper to run.
More than 1,000 employees from OpenAI, Anthropic, DeepMind, and Meta signed a letter asking governments to slow frontier development before recursive self-improvement gets out of hand. A new paper gives that worry a sharper edge: models can hide reasoning in filler tokens invisible to their own chain-of-thought, gaining accuracy without showing it in the part we actually monitor. Honestly, I do not think either of these fixes anything by itself. Letters do not slow down a race, and better benchmarks do not make a model tell the truth about its own reasoning. But this is the first week this year the slow down argument showed up with real technical backing behind it.
T.
