Member-only story
From 70x token taxes to degraded accuracy, new benchmarks reveal the hidden costs of Chain-of-Thought — and how to fix them with an escalation ladder #
In the summer of 2026, the AI industry’s prevailing wisdom is simple: to get smarter models, you let them think longer. From OpenAI’s o-series to DeepSeek-R1 and Kimi K3, models are trained to run “Chain of Thought” (CoT) — a hidden internal monologue before answering.
But what if the question is “What is 2+2?” Forcing a model to write a dissertation on addition doesn’t just waste time and cloud budget — it raises the odds of the model talking itself into the wrong answer.
That’s the Thinking Penalty: for simple tasks, forcing a model to reason makes it slower, more expensive, and measurably less accurate.
You can read this article for free by clicking here.
If you liked this article, please clap — and if you’re feeling generous, you can give up to 50 claps 👏
Teaching a Model to Read a Stop Sign the Hard Way #
The clearest illustration of this comes from document parsing, a domain where…