Anthropic restricts context editing and reasoning access on new API accounts to fight distillation

Anthropic says "preserved thinking," introduced with Claude Fable 5.1 and Claude Opus 5.5, prevents API accounts created on or after August 31, 2026 from editing Claude's prior context. With Claude Sonnet 5.5, reasoning transcripts are accessible only to the organization whose request produced them. Anthropic also says the reasoning from these models cannot be read by earlier models with weaker safeguards. A two-stage classifier now runs on Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1, with a stated false positive rate under 0.1%. Since September 24, Anthropic charges for requests its safeguards block before Claude responds, in areas where that rate is under 0.1%.

On results, Anthropic reports that reasoning tokens extracted by distillers on its platform are 10 to 100 times lower since the classifier launched. In an earlier test it reports, a model trained on summarized reasoning scored 71.8% on SWE-bench Verified against 80.4% when trained on full transcripts. These are Anthropic's own estimates on its own platform, and the company writes that it has "yet to see a highly sophisticated and persistent attempt" to get around the defenses. It also says that "preserved thinking has required some developers to update their integrations," without saying which integrations or what changed for them.

Why it matters: If your harness trims, rewrites or replays earlier turns, or passes reasoning between models or accounts, test that path on Fable 5.1 and Opus 5.5 with an API account created after August 31, 2026. The post does not list which workflows break, so treat this as a compatibility check rather than a known failure. Budget for blocked requests being billed.