Research note
Research note: Anthropic says new defences cut data harvested by distillers from Claude by around 90 to 99 percent
How we reported Anthropic’s distillation defences, why its figures are labelled company claims, and how we handled the companies it names.
What we set out to check
What Anthropic changed, what results it claims, who it accuses, and what it admits it does not know.
What we found
- Anthropic’s post describes preserved thinking, a classifier, account bans and billing for blocked requests.
- Its reduction figures are measured on its own platform.
Numbers that differ
The post’s summary says “more than 90%” while the body says “around 90 to 99 percent” and “10 to 100+ times”. These are consistent; we use the body’s wording.
Labels we chose and why
- Confirmed for the changes Anthropic made to its own products.
- Company claim for results and for its allegations about other companies.
- We note that we found no public responses from Zhipu or Moonshot, to be fair to them.
What we could not verify
- Independent measurement of the reduction.
- Responses from the companies named.
Disclosure
Anawords uses Claude as a research tool. We applied the same rules to this story as to any other.
What we searched
- Anthropic distillation defenses 90 99 percent
- Anthropic preserved thinking API accounts
- Zhipu GLM-5.3 distillation Claude
Every source we read
| Source | Used? | Why |
|---|---|---|
| Anthropic | USED | Primary source for the changes, figures, named companies and caveats. |
| AI Weekly | USED | Independent summary matching the key figures and noting they are unverified. |
| TechCrunch (September) | NOT USED | Earlier coverage of a separate Anthropic report; background only. |