Anthropic cuts Claude distillation by 90-99% with new reasoning-protection defenses
In short: Anthropic published details on how it curbed 'illicit distillation' — large-scale covert operations that scrape Claude's reasoning transcripts to train rival models cheaply. New measures, including reasoning summarization, 'preserved thinking' for new API accounts, and a two-stage classifier system, reportedly cut the volume of usable reasoning data extracted by distillers by 10-100x. Anthropic says a Chinese lab, Zhipu, used distillation (including attempts against Claude) to build GLM-5.3, which US assessors found to be the most cyber-capable open-weight model released so far with safeguards that are easily bypassed.
This summary was generated automatically by AI from Anthropic's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Reasoning summarized before being shown to users to make transcripts less useful for training rival models
- 2'Preserved thinking' prevents new API accounts (created after Aug 31, 2026) from editing Claude's prior context to extract reasoning
- 3Reasoning transcripts restricted to the organization that generated the request (Claude Sonnet 5.5)
- 4Two-stage classifier system (similar to bio/cyber safeguards) blocks extraction attempts with <0.1% false positive rate
- 5Accounts caught repeatedly attempting extraction face stricter limits or are banned, often within two hours
- 6Since Sept 24, requests blocked by these safeguards are billed as before response generation
| Parameter | Before | Now |
|---|---|---|
| Reasoning transcript access | Full chain-of-thought reasoning could often be extracted via active prompting or proxy accounts | Reasoning summarized, access restricted to originating organization, and new API accounts (post Aug 31, 2026) cannot edi |
| Distillation detection | Simple summarization-based defenses, bypassed by active extraction techniques | Two-stage classifier (constitutional classifiers approach) with <0.1% false positive rate, matching bio/cyber safeguard |
| Extracted reasoning tokens | Baseline high volume harvested by distillation networks | 10-100x lower (some days too low to measure) after full rollout |
Why it matters
Teams integrating with Claude's API — including those relying on chain-of-thought reasoning for debugging or chaining agent logic — may see changes in how reasoning is exposed, especially for accounts created after Aug 31, 2026. It also signals that reasoning-heavy model outputs are increasingly treated as a protected asset across the industry, which could affect transparency/debuggability of agent pipelines built on top of frontier reasoning models.
What it means for AI agents and contact centers
If your voice AI or automation stack calls Claude models (e.g., Opus, Sonnet, Fable) for reasoning-heavy tasks such as call analytics or complex tool-calling logic, check whether 'preserved thinking' restrictions on new API accounts affect your integration, and confirm you are not inadvertently triggering distillation-detection safeguards (e.g., by requesting verbatim reasoning output) which could lead to stricter rate limits or billing for blocked requests.
Sources
- AnthropicOfficialPrimary sourceOriginal article →„Illicit distillation allows model developers to copy frontier AI capabilities without creating proportionally robust safeguards. Here, we describe how we’ve reduced the rate of distillation by more than 90%.“12 Oct 2026, 03:00
- Published by source
- 12 Oct 2026, 03:00
- Found by our system
- 20 Oct 2026, 20:16
- Summary generated
- 20 Oct 2026, 20:18
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy