DecisionPivot is a diagnostic benchmark for decision models used in large language model (LLM) systems. It comprises 48 tasks and 40,033 items across seven decision roles. The core track holds 7,800 items generated from 13 task families and organized into contrast groups; each group's answer is verified by an independently implemented solver. The reference track adds 32,233 items drawn from 35 public test sets, preserving their original evaluation metrics.\ \ We evaluate 40 decision models and 14 general models using documented, model‑specific protocols, measuring item accuracy, contrast‑group accuracy, probability ordering, calibration, and empirical selective coverage. The results expose substantial differences hidden by aggregate scores: two 27B models differ by only 2.6 points on the reference track yet achieve contrast‑group accuracies of 70.1% and 44.2%. Across 2,400 binary contrast pairs, Perplexity Decider ranks the positive member above the negative one in 92.3% of cases but answers both correctly in only 74.1%, revealing a gap between probability ordering and decision correctness. Models with similar item accuracy also vary in the reliability of their high‑confidence predictions. On a 1,296‑item subset, GLM‑5.3 reaches 93.5% contrast‑group accuracy under a higher reasoning setting, compared with 81.5% at its lowest effort level. By linking errors to pivotal details and difficulty levels, DecisionPivot offers fine‑grained diagnostics of decision failures that aggregate evaluation alone cannot reveal.\ \ Review