Pin extended-thinking cost to zero on routine turns with MAX_THINKING_TOKENS=0

Claude Code Changelog v2.1.166 (Releasebot)·low signal

Claude Code v2.1.166 made MAX_THINKING_TOKENS=0 and --thinking disabled work reliably, so you can switch off extended reasoning on mechanical or low-stakes turns where it only adds latency and token spend. Reserve thinking budget for genuinely hard planning and debugging steps and disable it for boilerplate edits, formatting, and scripted runs. For unattended scheduled agents this is a direct, predictable lever on per-run cost.