AI news · October 13, 2026
Anthropic says new controls cut model-distillation data by 90% to 99%
Anthropic reports a 90% to 99% reduction in harvested data after its October 12 defenses; API teams should expect more account checks and less access to reusable reasoning traces.
- reported reduction in harvested data
- 90-99%
- cutoff for new API accounts
- Aug 31
- defense methods described
- 3
Anthropic says it has reduced the amount of Claude data collected by model distillers by around 90% to 99%. The company describes distillation as industrial-scale extraction through networks of fraudulent accounts, often aimed at copying a model's capabilities without the original developer's permission. Its new controls cover where reasoning can move, how accounts are linked and how suspicious usage is challenged. Anthropic says the measures are designed to limit attacks while keeping ordinary users productive.
One concrete change affects preserved thinking in Claude Fable 5.1 and Claude Opus 5.5: new API accounts created on or after August 31, 2026 cannot edit Claude's prior context in the way distillers used to. The company also says the controls have trade-offs, but gives no independent audit of the 90% to 99% figure. Builders should treat model outputs as service responses, not as a reliable training feed, and should design their own evals from permitted data.
What you can do with it
Read the defense note with your API and evaluation owners. Check whether your test harness depends on full reasoning traces, repeated account creation or replaying model context, then replace those assumptions with approved logs, labels and independent test cases. Re-run a small evaluation after any provider policy change.
Our take
This is a major change for teams that use frontier APIs to create training data or detailed evaluation traces. Anthropic's result is a company report, not an outside measurement, but the implementation detail is clear enough to plan around. Privacy, account stability and permitted data use now belong in the model selection decision.
Source: Anthropic — Made With Models writes the brief; the reporting is theirs.