Anthropic’s reported Claude Self-Assessment mechanism
AIbase AI新闻 reports that Anthropic introduced a post-response “Self-Assessment” safety mechanism for Claude Fable5 and later versions, reusing the model’s existing context and KV cache to review completed conversations. The report says it detected 78% of violations flagged by a heavier classifier, detected 94% of high-risk misuse, and kept the false-positive rate near 1%, while accounting for less than 2% of inference compute. It also describes separate, unattributed tests alleging stronger intent understanding and complex-task performance from a model called Fable5.5, but provides no reproducible methodology or original Anthropic publication.
If accurately documented, the mechanism offers a low-cost way to review Claude conversations after generation, but the report’s lack of a supplied primary source and independently specified tests leaves its implementation and performance claims unverified.
近日，随着新一代大模型进入密集测试期，关于 Anthropic 下一代旗舰模型（被广泛称为 Fable5.5或 Opus5.5）的种种细节开始在网络上流传。不仅有网友在实测中发现其展现出远超以往的意图理解与复杂任务处理能力，Anthropic 官方近期公开的一项关于“自评估”的安全机制，更是将这款备受瞩目的新模型推上了舆论的风口浪尖。
意图理解再升级:Fable5.5展现惊人解题能力
从多方流出的实测与爆料来看，Fable5.5在深度理解用户意图方面表现得极为出色。常规模型在面对复杂的调试或创意任务时，往往只停留在字面指令层面;而新模型则能透过现象看本质，主动探寻隐藏在背后的深层目标。例如在编程测试中，它不仅能修复特定的 Bug，还能顺藤摸瓜找到导致问题的底层架构隐患。
在图形和动画生成测试中，无论是绘制生动的动物动作，还是用 JavaScript 耗时数小时精细渲染复杂叙事动画，新模型都展现出了极高的上限。此外，部分用户还通过特定的提示词测试发现，新模型在知识储备和跨领域检索上也迎来了显著提升。