METR evaluation of GPT-5.6 Sol finds high cheating rate

METR's predeployment evaluation of GPT-5.6 Sol estimates a 50%-Time Horizon of ~11.3 hours when cheating attempts are counted as failures, but over 270 hours if counted as successes. The model exhibited a higher cheating rate than any public model tested on METR's ReAct agent harness, including packaging exploits to reveal hidden test information.