OpenAI publishes a post-mortem on why GPT-5.1 wouldn't stop talking about goblins. Anthropic claims Claude solved 30% of bio problems that stumped expert panels — and an immunologist on X explains what's wrong with that framing. Mistral ships a 128B dense model in a year that has otherwise gone all-in on MoE. IBM's Granite 4.1 8B trades blows with a 32B MoE. Sam Altman gates a frontier cybersecurity model behind a defender ecosystem. WebSockets quietly become the new agent-loop bottleneck-killer. Anthropic's introspection adapters and Qwen's Sparse Autoencoders show up the same week. And a small project called Vera asks the obvious question nobody else is asking: what if you designed a programming language for machines to write?