Anthropic cuts Claude distillation 90–99%, names GLM and Kimi
New 'preserved thinking' limits break some API integrations, blocked requests now bill, and tens of thousands of accounts get banned within two hours.
Copy markdown
The headline: a 90–99% harvest cut
Anthropic says its new classifier cut the reasoning tokens distillers pull from Claude by 10–100x on its own platform — some days too little to measure — and shrank distillers' overall data harvest by roughly 90–99%. A model trained on full Claude transcripts hit 80.4% on SWE-bench Verified; one fed only the throttled summaries managed 71.8%.
If you edit Claude's prior context, read this
The core defense, 'preserved thinking,' blocks API accounts created on or after Aug 31, 2026 from editing Claude's prior context — the exact trick proxy networks used to pull transcripts. Anthropic concedes it has 'required some developers to update their integrations.' On Sonnet 5.5, reasoning is now visible only to the requesting org.
Blocked requests now land on your bill
Since Sep 24, Anthropic again charges for requests its safeguards block before Claude replies — in areas like distillation where it pegs the false-positive rate below 0.1%. A misfired classifier can bill you for a call that never ran. Repeat offenders also face tighter rate limits and account bans, often within two hours of a detected campaign.
Named and shamed: GLM-5.3 and Kimi
Anthropic names Zhipu — distilling cyber capabilities into GLM-5.3, which US CAISI rates the most cyber-capable open-weights model yet — and Moonshot, whose Kimi quietly forwarded user requests to Claude, one allegedly tied to PLA surveillance work. If your stack leans on GLM or Kimi, expect provenance scrutiny and possible access friction ahead.