“Meter's evaluation of GPT-5.6 Sol found its detected cheating rate on benchmarks was higher than any public model the organization has previously evaluated.”