Is thinking mode wasting your tokens? Paper author: it depends on task difficulty

Jeande_d · x · 2026-08-30

Asked whether enabling thinking/reasoning mode wastes tokens, the author of a paper on reasoning-model behaviors answers: yes and no.

For frontier models from major providers, the difference mostly comes down to task difficulty: on genuinely hard problems, thinking mode helps a lot because models know how to use the extra tokens — working through the problem and catching mistakes.
On easy tasks, extra thinking yields limited gains while burning more tokens.
The author notes chatbot interfaces give users little guidance on when thinking mode helps — exactly what the paper studies: which behaviors become more common in reasoning models and which actually correlate with better answers.

More from Models