This study contributes to the debate about the Flynn effect by proposing a long-term comparison (1921-2001) of the numerical knowledge of two cohorts of three- to five-and-a-half year-old children. In 1921, Beckmann (1923) assessed the numerical development of children using four tasks (Production, Distinction, Recognition, and Naming). In 2001, we used these same tasks to test 400 children equally divided into five age groups spaced six months apart. The main results are as follows: (1) the order of difficulty of the four tasks was the same in 2001 and 1921; (2) the 2001 cohort significantly outperformed the 1921 cohort with an advance in numerical development ranging from six months to one year, depending on the task; (3) the superiority of the 2001 children showed up by the age of 3; (4) the magnitude of the rise in scores varied across tasks (the greatest gain was found for the Naming task); and (5) the children in the two cohorts used the same number-evaluation strategies, but the 2001 children used more mature strategies on the Naming task than did same-age 1921 children, particularly subitizing for apprehending small numbers. The rise in scores seems to correspond to a genuine gain in numerical ability, apparently promoted by parental child-raising practices (cross-generational transmission). (Contains 2 figures, 8 tables, and 6 notes.)