THE HOOK
Anthropic has gone from complaining about copycats to naming them, and says it has slammed the door on most of the free lunch.
THE GIST
Anthropic says its new defences cut Claude reasoning tokens extracted by distillers 10 to 100 times. It combines preserved thinking, an extraction classifier and fast bans, with tens of thousands of accounts banned. It alleges Zhipu and Moonshot ran distillation campaigns.
THE TAKE
Good, because distillation at industrial scale is just piracy with a GPU bill. But a classifier that bans accounts in two hours will catch honest users too, and Anthropic now charges for some blocked requests. Security theatre it isn't; collateral damage it is.
OPINION
The Decoder reports Anthropic admits legitimate users are getting caught up. The national security angle, including an alleged PLA-linked surveillance request routed through Kimi, all but guarantees this ends up in Washington's export-control debate.
- Story type: Important story
- Impact: Medium
- Confidence: High (official announcement)
- Audience: Developers
- What changed: Anthropic disclosed anti-distillation defences that it says cut extracted Claude reasoning by 90 to 99% and named Zhipu and Moonshot.
- What to watch: Whether other labs adopt similar defences, and whether US officials act on the Zhipu and Moonshot allegations.
- Reading time: 1 minute
