Anthropic says new defences cut data harvested by distillers from Claude by around 90 to 99 percent
The company says it blocks new accounts from editing Claude’s earlier reasoning and has banned tens of thousands of accounts. It names Zhipu and Moonshot; the figures are its own.
How this story was checked
- ResearchedWith AI assistance · 2 sources
- AI source checkEach sentence matched to its source
- Editor approvalMayank Sahu · 13 Oct, 16:33
- Published13 Oct, 16:33
In short
Anthropic said on 12 October 2026 that new defences against distillation, where others copy a model by training on its outputs, have cut the data harvested from Claude by around 90 to 99 percent. The figures are Anthropic’s own and not independently checked.
What Anthropic announced
Anthropic said on 12 October 2026 that new defences against “distillation” have reduced the data harvested from Claude by “around 90 to 99 percent”, based on early evidence.12 Distillation means training a new model on another model’s outputs to copy its abilities.
Anthropic says its measured extraction of reasoning text fell 10 to 100 times or more after a detection system was fully deployed.1 These figures come from Anthropic’s own platform and have not been checked by anyone else.12
What changed
- Preserved thinking: with Claude Fable 5.1 and Claude Opus 5.5, new API accounts created on or after 31 August 2026 can no longer edit Claude’s earlier context, a method Anthropic says was used to pull out long reasoning transcripts.12
- Detection: a two-stage classifier runs on several Claude models, with a false-positive rate below 0.1%, Anthropic says.12
- Enforcement: Anthropic says it has banned tens of thousands of accounts linked to distillation, often within two hours of a campaign starting, and since 24 September bills for some blocked requests.12
Who it names
Anthropic says Zhipu distilled cyber capabilities while developing GLM-5.3, and that requests to Moonshot’s Kimi models were secretly forwarded to Claude.12 These are Anthropic’s allegations; we did not find public responses from the companies named when we checked.
Why reasoning matters
Anthropic says a model trained on Claude’s full reasoning scored 80.4% on the SWE-bench Verified coding test, against 71.8% when trained on summaries only, which is why it protects full reasoning text.1
Limits Anthropic admits
It calls this a “cat-and-mouse game”, has not yet faced a highly sophisticated attempt, says some legitimate users may be blocked by mistake, and says some developers must update their code.1
Disclosure: Anawords uses Claude, made by Anthropic, as one of its research tools. Anthropic has no say in our coverage.
Fact ledger
How labels workQuestions readers ask
What is AI distillation?
Training a new model on another model’s outputs to copy its abilities, often without permission.
What did Anthropic change?
It blocks new API accounts from editing Claude’s earlier reasoning, runs a classifier to detect distillation, and bans linked accounts.
Are the 90–99% figures independent?
No. They are Anthropic’s own measurements.