❯ Anthropic reports a 90–99% reduction in extracted reasoning data, with integration trade-offs for developers

The measure is extracted dataIn an October 12 technical post, Anthropic estimated that new defenses reduced the Claude reasoning tokens extracted on its platform by approximately 90–99%. This measures data extraction on the company’s own service, rather than an equivalent decline in other models’ capabilities or the elimination of all distillation.

Reasoning movement becomes more constrainedMeasures include preserved thinking in Fable 5.1 and Opus 5.5, restrictions on new API accounts editing prior context, and controls on protected reasoning moving between models, accounts or organizations. A two-stage classifier identifies extraction requests, with stricter limits for persistent attempts. The focus is on protecting the reasoning process.

Legitimate users also face costsAnthropic acknowledges that some developers must change integrations, normal requests can be blocked incorrectly, and blocked requests in relevant low-false-positive categories are charged. Enterprise users need to review context handling and billing. Protecting research and preserving API usability involves concrete trade-offs that one headline percentage cannot capture.

▮ SIGNALModel vendors are treating reasoning data as a managed asset, with protections that directly affect developers’ context design, costs and user experience.