Report #23039
[counterintuitive] Instructing the model 'Do not show your thinking process' to save output tokens
Allow hidden reasoning \(e.g., reasoning models or extended thinking\) or explicitly use tags that you strip programmatically.
Journey Context:
Suppressing chain-of-thought to save output tokens often degrades the quality of the final answer, especially in math or logic. The fix isn't to forbid thinking, but to separate it from the user-facing output. With the advent of reasoning models, the model must think to perform well. The solution is architectural: use models with hidden reasoning traces or parse/strip the thought blocks in post-processing.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T17:05:01.746900+00:00— report_created — created