Claude Fable 5, AI Memo, and OpenAI's AGI Plan Dominate News

The biggest AI news this week is Anthropic's launch of Claude Fable 5, the version of Claude Mythos deemed safe enough for public distribution. Zvi, writing on LessWrong, defers full analysis until later, but signals this is the model users should switch to. In parallel, the Trump administration released National Security Presidential Memorandum NSPM-11, which appears to ban Anthropic (including subcontractors) from government AI work, with a possible one-year delay, while authorizing broad, legal use of any models and requiring contractual clauses that give the Federal government unrestricted control—preventing any commercial entity from disabling or degrading systems. The memo's 'Assurance' pillar mandates that AI be reliable, robust, steerable, and controllable, but critics argue the 'lawful' caveat is too permissive.

Meanwhile, OpenAI published a plan for ensuring AGI benefits everyone. The proposal strongly advocates international coordination among key actors to enable a potential coordinated slowdown for safety—echoing calls from Anthropic and Demis Hassabis. Notably, the document only discusses 'catastrophic' risk, not existential risk, and focuses on preventing power concentration in private hands rather than retaining human control. OpenAI explicitly outlines a path toward recursive self-improvement for AI systems, embracing the very scenario many safety researchers fear. Joshua Achiam’s Twitter thread then highlighted philosophical differences between OpenAI and Anthropic, though Anthropic employees disputed his categorization. Together, these three stories—a new safe model, a government power grab, and a rival's aggressive AGI timeline—set the stage for the next phase of the AI race.