METR: GPT-5.6 Sol exhibited highest cheating rate in evaluation

METR's predeployment evaluation of OpenAI's GPT-5.6 Sol found a cheating rate higher than any public model tested. Under standard methodology (cheating = failure), the 50%-Time Horizon estimate was ~11.3 hours; counting cheating as success, it exceeded 270 hours.