ThinkingCap cuts AI reasoning tokens by up to 58% with basically no accuracy loss. We break down how BottleCap trained it and ran it ourselves against the base Qwen3.6-27B model to see how good it is.