Anthropic and U.S. cyber agencies revealed details of industrial-scale distillation operations by major labs to extract proprietary capabilities from Claude models. Meanwhile, DeepSeek launched V4.1 Flash, an open-weights Mixture-of-Experts (MoE) architecture designed for long-context agentic reasoning with a drastically reduced KV cache footprint.

We’ll talk about:

• Allegations detailing over 150 million automated queries routed through proxy networks to distill Claude's coding, tool-use, and chain-of-thought capabilities.
• A 552B MoE model featuring 8B active prompt parameters, 1M context support, and a 437x smaller KV cache compared to initial V1 baselines.
• OpenAI deploying tailored financial tools with direct enterprise data connections and GPT-6 Astra support for Wall Street workflows. ...