The "thinking" feature detection has been improved to use proper word boundary regex patterns instead of simple string includes. This prevents false positives when these keywords appear as part of other words.
Thinking triggers (with proper word boundaries):
- "think" → 4,000 tokens
- "think hard", "think deeply", "megathink" → 10,000 tokens
- "think harder", "ultrathink" → 31,999 tokens