AI agents make poor long-term decisions, new Taste-Bench benchmark finds

Researchers introduce Taste-Bench, a benchmark of 502 binary decision-fork questions that tests whether AI agents pick the better long-term path. The best of 14 tested models, GPT-5.6 Sol, scored 59.7 percent, against 25 percent for random guessing, and accuracy fell to 21 percent when the deciding evidence appeared only later in a run.