Anthropic says it's making it harder for rivals to clone Claude's capabilities on the cheap.
The company published research outlining new technical defenses against distillation attacks. In a distillation attack, an outside party queries a model at scale, logs its outputs, and uses them as training data for a smaller, cheaper model that mimics the original's behavior. Anthropic frames the work as protecting the engineering and compute investment behind its frontier models, though it hasn't detailed exactly which countermeasures it's deploying in production.
This matters because distillation is already how several lower-cost competitors have closed the gap with pricier frontier labs - DeepSeek drew scrutiny last year over allegations it trained on outputs from OpenAI's models. If Anthropic can make that kind of copying meaningfully harder, it buys more time for the capability lead that its subscription and API pricing depend on.
Still, "defenses against distillation" is also a polite way of saying "stop using our model to build a cheaper one" - and no outside lab has yet tested whether these measures hold up against people actually trying to get around them.