Dev Uses Jev to Dynamically Tune GPT-6 Reasoning Effort Inside Codex, Cutting Costs 50%

brandon_galang · x · 2026-09-21

Developer @miu21590 shows a novel use of Jev: instead of picking a model before a task, Jev adjusts GPT-6's reasoning effort inside Codex during the task — more thinking when stuck, less for routine steps. In their tests this cut Astra costs by 50%, ran faster, and preserved prompt caching.

Original post →

More from coding & agent

coding & agent channel →