Anthropic says it has cut data harvesting from Claude by up to 99%

Expand

Open original image ↗

Anthropic says its new safeguards have reduced the reasoning harvested from Claude by around 90–99%. Its 12 October update measures extraction on Anthropic’s own platform; it does not establish an equivalent drop across third-party providers.

The target is unauthorized distillation: collecting a model’s responses to train another model. Detailed reasoning is particularly valuable because it supplies the steps behind an answer.

A sharp drop in extracted reasoning

Expand

Open original image ↗

Anthropic combines restrictions on moving reasoning between contexts with extraction detectors, tighter usage limits and account bans. It says legitimate requests are wrongly flagged in fewer than 0.1% of cases. Some blocked requests are now charged, and the company acknowledges that developers have had to change integrations.

What changes for developers

The practical issue is preserved thinking. According to the API documentation, protected reasoning depends on the conversation that produced it. Rewriting earlier messages, changing tools or rebuilding a system prompt can invalidate thinking blocks sent back on a later turn.

For supported models, accounts created on or after 31 August 2026 face this conversation check by default. Older accounts have different enforcement rules. Anthropic recommends keeping earlier context intact and appending new information, so integrations behave consistently across accounts.

Claude Code and claude.ai users do not need to rebuild their workflows for this check. Developers with custom history management are more likely to be affected. Dropping invalid thinking can keep requests moving, but the documentation warns that recreating lost reasoning can increase output usage and response time.

The reported reduction is substantial, but its durability remains unproven. Anthropic says the defenses have yet to face a highly sophisticated, persistent attempt to defeat them.