Anthropic's distillation numbers, labeled as their own test

On October 12, 2026, Anthropic published early evidence on distillation defenses. The post says those defenses left 90–99% less harvested data. Nobody outside Anthropic has published a check of that range. The figures below are the ones Anthropic printed.

What they measured

On SWE-bench Verified, Anthropic says a pretrained Claude trained on full Opus 4.8 transcripts scored 80.4%, against 71.8% when the same kind of model was trained on summarized transcripts. That is their test of transcript value, not a public leaderboard Littlewolf ran.

They also say reasoning-token extraction is 10–100x lower, and that a two-stage classifier on Sonnet 5.5, Opus 5, Opus 5.5, Fable 5, and Fable 5.1 has a false positive rate under 0.1%. They say they have been charging for blocked requests since September 24 where that false-positive rate holds.

What they say about accounts

Anthropic says tens of thousands of accounts were banned, often in under two hours. The post names Zhipu's GLM-5.3, which it says CAISSI assessed as the most cyber-capable open-weights model, plus Moonshot / Kimi, and a request it describes as affiliated with PLA-linked CCTV. Those are Anthropic's characterizations of the incidents it describes.

A separate control in the same write-up: preserved thinking. New API accounts created on or after August 31, 2026 cannot edit prior context.

What this is not

This is not a Littlewolf benchmark. We did not harvest transcripts, rerun SWE-bench, or measure the false-positive rate. If you need a number you can reproduce yourself, it is not on this page.