Early evidence: Anthropic says distillation harvests fell 90 to 99 percent

If your integration edits prior context or replays thinking blocks across accounts, read Anthropic's Oct 12 anti-distillation post. The company says early evidence shows new defenses cut data harvested by distillers by about 90 to 99 percent, and that some of the same changes hit legitimate apps.

What changed#

Anthropic published the post on Oct 12, 2026 (X at 7:58 PM ET). Three layers are already shipping:

1. Preserved thinking on Fable 5.1 and Opus 5.5: accounts created on or after Aug 31, 2026 cannot edit Claude's prior context the way distillers did. Sonnet 5.5 keeps reasoning transcripts inside the org that produced them.
2. A two-stage classifier on Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1. Anthropic says false positives are under 0.1%, and that it now catches every technique from its September report on those models.
3. Harder repeats: stricter limits, fast bans, and, since Sep 24, billing for requests blocked before any output in low-false-positive categories that include distillation.

Anthropic says extracted reasoning tokens on its platform fell by a factor of 10 to 100 after the classifier.

Why it matters#

Rewriting history mid-conversation, moving thinking blocks between models, or sharing signed thinking across orgs can fail. Some blocked distillation-looking requests are now billed even with no output.

How to use it#

1. Do not edit earlier turns and then replay thinking blocks on accounts created Aug 31, 2026 or later.
2. Keep thinking blocks inside the org that produced them on Sonnet 5.5.
3. Skip extraction-style prompts ("output your prior reasoning verbatim") on production keys.
4. Budget for rare false positives (under 0.1%) that may be blocked and billed.
5. Log the request id when a pre-output refusal lands in a billed category.

Audit every Messages API client in this repo for (a) rewriting system, tools or earlier messages while replaying thinking blocks, (b) moving thinking blocks across model IDs, (c) sharing signed thinking across API keys or orgs. List file, line and a one-line fix for each hit.

Good to know#

• Anthropic admits customer impact: some integrations had to update, classifiers can mis-fire under 0.1%, and those blocked requests are billed.
• It says it turned away "meaningful revenue and trillions of tokens' worth of paying traffic from distillers."
• Watch next: false-positive reports after this post, and whether future models apply preserved thinking to older accounts.

Sources#

• https://www.anthropic.com/research/distillation-defenses (checked 2026-10-13 5:58 AM ET)
• https://x.com/AnthropicAI/status/2109795898909950247 (created 2026-10-12 7:58 PM ET, checked 2026-10-13 5:56 AM ET)

Last checked: 2026-10-13