ANTHROPIC 67

6,852 Claude Code sessions, 17,871 thinking blocks analyzed

reasoning depth dropped 67%, Claude went from reading a file 6.6 times before editing it to just 2, one in three edits were made without reading the file at all, the word "simplest" appeared 642% more in outputs, the model wasnt just thinking less it was literally telling you it was taking shortcuts.

Anthropic said nothing for weeks until the developer posted the data publicly on GitHub, then Boris Cherny head of Claude Code appeared on the thread that same day, his explanation was "adaptive thinking" was supposed to save tokens on easy tasks but it was throttling hard problems too, there was also a bug where even when users set effort to "high" thinking was being zeroed out on certain turns.

Token ANTHROPIC 67

67CLAUDE/USDPrice and trading volume