Claude Fable 5 Spills Raw Inner Reasoning in Web Interface

Anthropic's top-tier model "Claude Fable 5" was reported in early July 2026 to have output its normally hidden internal reasoning (Chain-of-Thought) directly in its web interface. The trigger came when a user fed Fable 5 a hard Codeforces problem (contest 2237/problem/H and a simplified variant) and, instead of a clean answer, the model displayed its raw reasoning trace. The content was fragmentary, mixing onomatopoeia and abbreviations like "DATA DATA DATA. GO.", "GRRR", "GAAAH" and "PHEW" with graph-theory and competitive-programming notation and a 💀 symbol marking dead ends, suggesting the model reasons in an efficient internal representation it devised on its own. The phenomenon was shared in a Reddit thread and drew wide attention.