Jev Experiment Dynamically Tunes GPT-6 Astra Reasoning in Codex to Cut Reported Costs

A developer experiment is testing whether Jev can act as a real-time reasoning-effort controller for GPT-6 Astra inside Codex, raising effort when an agent is stuck and lowering it for routine steps. The unpublished test claims roughly 50% lower Astra costs while preserving prompt caching, a result that depends on GPT-6 Astra's cache-preserving configuration_update mechanism rather than ordinary request-level effort changes.