Anthropic Cuts Claude Reasoning-Token Theft by Up to 99%

What happened

Anthropic says it has sharply reduced the theft of Claude's internal reasoning tokens, reporting cuts of between 90 and 99 percent following a crackdown on what it describes as industrial-scale distillation attempts. The company has banned tens of thousands of accounts linked to the activity.

Anthropic attributes a significant share of the large-scale extraction attempts to Chinese AI companies Zhipu and Moonshot. Reasoning tokens are the intermediate "thinking" steps a model produces before delivering a final answer; harvesting them at scale allows a third party to train or fine-tune a rival model on another provider's reasoning patterns without replicating the underlying research.

The intervention appears to be a targeted enforcement and detection effort rather than a single product change, aimed specifically at accounts attempting to systematically pull reasoning traces out of Claude via the API.

Why it matters

Reasoning tokens have become some of the most commercially sensitive output a frontier AI lab produces, since they can encode the "how" behind a model's problem-solving, not just the "what." Protecting them is now a competitive and security priority on a par with protecting model weights themselves, because distillation lets a competitor approximate years of research spend relatively cheaply.

For enterprises and public-sector bodies building on top of foundation models, this signals that usage terms, monitoring and account-level enforcement are becoming as important to vendor selection as raw model capability. It also points to a widening front in AI competition: not just who trains the best model, but who can defend the intellectual property generated at inference time.

By the numbers

• 90 to 99 percent reduction in reasoning-token extraction reported by Anthropic.
• Tens of thousands of accounts banned by Anthropic in connection with the activity.
• Two named Chinese AI companies — Zhipu and Moonshot — identified by Anthropic as linked to large-scale distillation attempts.

The Renascence take

The headline figure is dramatic, but the more interesting story is what it reveals about how value is now concentrated in AI products — not just in the model weights, but in the live reasoning trace a user never explicitly asks for yet pays for implicitly.

Most organisations buying AI capability still think in terms of "which model is smartest," when the more durable question is "whose reasoning is defensible." If a lab's intermediate thinking can be siphoned off and repurposed by a rival at a fraction of the cost, the real moat shifts to detection, account governance and usage-policy enforcement — the unglamorous operational layer, not the model card. Enterprises should treat API terms, anomaly monitoring and account-level controls as a genuine part of their AI vendor due diligence, not boilerplate to skim past.