The defenses target Claude's reasoning transcripts, the most useful training data for copying a model. A two-stage classifier, built like Anthropic's bioweapons safeguards, has a false positive rate under 0.1% and runs on Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1. Preserved thinking stops API accounts created on or after August 31, 2026 from editing Claude's earlier context, which blocks the proxy networks distillers used to extract reasoning repeatedly. Reasoning tokens reaching distillers fell 10 to 100 times, and Anthropic resumed charging for requests its safeguards block in areas with under 0.1% false positives. Anthropic names Zhipu's GLM-5.3 and a Moonshot Kimi routing case as examples of distillation from Claude. It says it has turned away meaningful revenue and trillions of tokens of paying traffic, and that some legitimate users had to update integrations. For founders, two practical points. Teams that build on Claude and manipulate prior reasoning blocks may need to change their code. Anyone training on a frontier model's outputs now faces account bans, stricter limits and charges for blocked requests.