Taste-Bench: top AI models pick the better path in long tasks only 59.7% of the time

Taste-Bench, a benchmark from researchers at City University of Hong Kong and Microsoft first posted to arXiv on 22 September, tests whether AI agents make good choices partway through long coding and research tasks. Each of its 502 questions shows the work so far and two possible next directions, one of which later turned out better. A question counts only if the model picks correctly with the options in both orders, so random guessing scores 25%. The best of 14 models, GPT-5.6 Sol, scored 59.7%, with GPT-5.5 at 59.5%, and accuracy across models fell to 21.0% when the deciding evidence came much later. Giving models more reasoning time did not help. These are the authors' results.