⚡️ JUST IN: OpenAI says some previously released GPT-5 models were accidentally exposed to limited chain-of-thought grading during RL training, but found no clear evidence of major monitorability degradation.
$DYM $PLUME $ARB