What's wrong with this entry?
Anonymous. No account, no email.
What
A new system-level optimization for Opus 4.6 that tunes the model's extended thinking behavior, reducing unnecessary reasoning overhead on simple follow-up messages while preserving full thinking on complex tasks.
Details
- Adds a
thinking_guidancesection to the system prompt instructing Opus 4.6 to calibrate thinking frequency based on task complexity - On follow-up turns (after the first assistant response), appends a
<system-reminder>tag hinting the model to skip thinking blocks for straightforward actions - Only activates when: using Opus 4.6, thinking is not disabled, and no custom system prompt is set
- Gated by server-side flag
loud_sugary_rock— rolling out gradually
This should result in noticeably faster responses on simple follow-up requests (e.g., "run the tests", "fix that typo") while maintaining full reasoning depth for complex tasks that need it.
Evidence
Thinking frequency tuning system (search for "tune your thinking frequency") — gated by loud_sugary_rock in clientDataCache, system reminder at meK (search for "without a thinking block")
Strings lifted out of the shipped bundle, so the claim above can be checked against them.