Following Gemini 4 Carbon's release, Anthropic is testing its new model, Fable 5.5. Some users have been routed to this new model while using Opus 5.5, and hands-on tests show that it outperforms existing models such as Opus—it better understands users' deeper intentions, breaks free from conventional thinking, and produces better content. In addition, Anthropic has introduced a safety monitoring method called self-evaluation, which leverages the KV Cache reuse mechanism of large-model inference to have the model that has just finished answering—after the user's Q&A session ends—generate a safety review invisible to the user, with less than 2% additional compute overhead, screening conversations for potential abuse risks. This mechanism does not participate in real-time interception; the results are used only for offline aggregate analysis, to investigate missed risks, and to discover new types of risks. The associated compute costs are borne by the platform and do not affect the user experience or billing.