Anthropic Says New Defences Cut Claude Model Extraction By Up To 99%

Anthropic says it has made it substantially harder for rivals and other outside operators to extract the reasoning behind Claude’s answers for use in their own models. In research published on October 12, the company says new defences cut the number of reasoning tokens obtained by suspected model distillers on its platform by roughly 10 to 100 times.

Put another way, Anthropic estimates a reduction of about 90% to 99% in the particular material it measured. That is a company-reported result from its own platform, not proof that every attempt to copy a model can be stopped. Anthropic says it is still working against more persistent and sophisticated efforts.

The issue is becoming more consequential as advanced models expose reasoning traces that can teach another system how to solve difficult problems. A developer can ask a powerful model thousands of questions, collect the responses and train a smaller model to imitate them. That training method is called distillation. It has legitimate uses, including helping companies make their own models cheaper and faster. Anthropic is targeting unauthorized extraction from Claude, not arguing that all distillation is improper.

Its approach combines changes to how Claude reveals its thinking, systems that identify suspicious use patterns and action against accounts it believes are involved in abuse. The company points to preserved thinking in Claude Fable 5.1 and Opus 5.5, alongside classifiers and fraud controls. The aim is to reduce the useful training signal an extractor can collect while leaving ordinary developer applications workable.

That balance is not simple. Developers also need reliable outputs and predictable access, so a defence that incorrectly flags normal usage can impose costs of its own. Anthropic acknowledges the possibility of false positives and integration friction. Independent evidence of how the measures perform against determined attackers is still limited.

The research lands as AI companies are devoting more attention to both misuse and leakage. TechBooky has reported on hidden reasoning traces escaping from frontier models and on Anthropic’s earlier efforts to disrupt Claude misuse. Model extraction is a different problem: the target is not a user’s account, but the expensive capabilities a lab has spent money to build.

Anthropic’s numbers are striking, but the commercial test is whether the defences remain effective without degrading Claude for legitimate customers. As reasoning models become more valuable, guarding their outputs may become as important to AI labs as guarding their training data and model weights.