I sent Claude Opus 5 '–-' and it wrote me 5k tokens about a cartographer (austinsnerdythings.com)

Recent experiments have revealed an intriguing behavior in Anthropic's Opus 5 AI model, where sending a specific sequence of three unspaced dashes ('---') can trigger the model to produce lengthy documents—sometimes exceeding 5,000 tokens—despite a lack of context or prompt. This was notably observed by various users, leading to a deeper investigation into why such input causes the model to interpret it as an unfinished document requiring completion. The research involved 649 API calls to the Bedrock platform and showed variance in output length and content, indicating a peculiar mode selector responds to the structured input rather than viewing it as an incomplete message. This anomaly is significant as it highlights potential vulnerabilities in AI understanding of markdown grammar, revealing that Opus 5 behaves more like a text processor than a strict conversational agent when encountering certain formatting cues. The results also point to an output pattern focused on representations of absence, particularly through titles often related to cartography, demonstrating how AI can generate coherent narratives that reflect conceptual themes derived from the structure of the input. Such observations inform the AI community about the nuances in model design and operational context, urging developers to reconsider their approaches to prompt engineering and the handling of unexpected input formats.