Preserved thinking
What it is
Preserved thinking is an anti-distillation safeguard Anthropic ships in the Claude API: a model's "thinking" blocks (its visible chain-of-reasoning trace) are only replayable alongside the exact system prompt, tools, and preceding messages that produced them. If a developer edits that prior context and then tries to resubmit or reuse the thinking block, the API rejects the request instead of returning it. The mechanism doesn't hide reasoning — it makes the reasoning trace cryptographically tied to its original context, so it can't be silently detached and repurposed as training data.
Why it's emerging now
Distillation attacks — using large numbers of accounts to harvest a frontier model's outputs and reasoning traces at industrial scale, then training a cheaper copycat model on them — have become a live concern as reasoning traces themselves (not just final answers) turn out to be the most valuable and copyable asset a reasoning model produces. Anthropic introduced preserved thinking with Claude Fable 5.1 on September 1, 2026, framing it explicitly as a defense against this kind of theft, and extended the same safeguard to Claude Opus 5.5 on September 22, 2026 — which is what pulled a fresh round of independent coverage.
How it relates to what you know
It's the defensive counterpart to distillation, already on your ledger as a training technique: where distillation compresses a large model's behavior into a smaller one, preserved thinking is a specific mechanism for stopping that compression when it's done without authorization, by denying attackers clean access to reasoning traces. It sits in the same family as the broader "anti-distillation" research area (output-perturbation defenses, distillation fingerprinting) but is notable as a concrete, shipped API-level implementation rather than a research proposal.