Claude Tests Account-Level Reasoning Summaries

This looks more like observability and anti-distillation tuning than a capability regression, but account-level experimentation makes the change confusing for developers.

• –User reports suggest Anthropic is testing different summary behavior across accounts and model surfaces.
• –Anthropic says summarized reasoning preserves model capability while reducing misuse and transcript extraction risk. [Anthropic threat report](https://www.anthropic.com/threat-intelligence-report-september-2026)
• –Better summaries could make traces more useful for debugging instead of exposing noisy, low-signal internal narration.
• –Developers relying on visible thinking for agent monitoring should treat it as an unstable interface, not a guaranteed diagnostic signal.
• –The lack of an official rollout note leaves users guessing whether changes are intentional experiments, quality fixes, or capacity management.