Anthropic distillation defenses: what changes in Claude's API

What we know

• What distillation is: covert, large-scale campaigns to extract a model's capabilities and train another without permission, often through networks of fraudulent accounts, according to Anthropic. Claude's reasoning transcripts are the most sought-after data.
• The claimed result: 90 to 99 percent less data harvested, based on what Anthropic calls early evidence. It estimates the reasoning tokens distillers extract on its platform are 10 to 100 times lower than before. That is Anthropic's own estimate.
• Preserved thinking: introduced with Claude Fable 5.1 and Claude Opus 5.5, according to the report. The platform documentation adds that the check also applies to Sonnet 5.5 and Haiku 5.5.
• Who it hits by default: API accounts created on or after August 31, 2026, 00:00 UTC, according to the docs. Older accounts only if they opt in.
• What breaks it: editing, reordering or deleting earlier messages, changing the system prompt or the tools, or shortening tool results already sent. By default the API returns a 400 error.
• A new classifier: it targets attempts to extract reasoning. Anthropic puts its false positive rate below 0.1 percent. It runs on Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1.
• Blocked requests are billed: since September 24, Anthropic charges for requests its safeguards block before Claude responds, in areas that include distillation.
• Penalties: accounts that keep trying to extract reasoning face stricter usage limits and may be banned. Anthropic says it has banned tens of thousands.

What changes and what doesn't

For normal use, most requests behave as before. What changes is how you send a conversation's history when the model reasons: it has to go back exactly as it arrived, and you can only append new messages at the end.

Anthropic admits preserved thinking has forced some developers to update their integrations, and that it has not fully met its goal of leaving legitimate users untouched.

There is a stopgap. The "drop_block" option discards the affected reasoning and lets the request through. Anthropic calls it a stopgap, and it costs tokens: in its benchmark, using it once per session raised output tokens by 2.5 to 20.1 percent; on every turn, by 4.6 to 67.1 percent.

How to tell if you're affected and what to do today

1. Check which model your chatbot or automation uses. If it's Fable 5.1, Opus 5.5, Sonnet 5.5 or Haiku 5.5 with reasoning on, read on.
2. Check when your API account was created. On or after August 31, 2026, the rule already applies.
3. Look for the habits that break it. Trimming old messages to save tokens, injecting the date or user data into the system prompt on every turn, or changing the tool list mid-conversation. The docs offer alternatives, such as in-conversation system messages and server-side compaction.
4. If you rely on a plugin or a third-party service, ask the vendor whether it passes reasoning blocks back untouched and whether it has tested these models.
5. Check your error logs. A 400 "invalid_request_error" that started when you switched models points here. Use "drop_block" only while you fix the cause.
6. Watch model switches. If a router moves a conversation from Fable 5.1 to any model other than Mythos 5.1, that reasoning is lost, according to the docs.
7. Don't ask Claude to repeat its reasoning word for word. It's one of the techniques the classifier looks for, and a blocked request is billed.

Related: Claude Haiku 5.5: Up to 90% Cheaper, With a 100K-Token Catch

Sources

• Anthropic, "Strengthening our defenses against distillation attacks," October 12, 2026
• Anthropic, Claude platform documentation, "Preserved thinking"

Updates: if Anthropic changes the account cutoff date, the models covered or the billing of blocked requests, it will be added here with a link.