Study 1 investigated whether the strength of correlation between latent variables representing working memory capacity (WMC) and fluid intelligence (Gf) depends on the time allowed to work on an intelligence test. When the half recommended time was given to fulfill two Gf tests, WMC and Gf were statistically indistinguishable, indicating that working memory and fluid intelligence are fully isomorphic constructs. However, when virtually no time limit was applied, WMC explained only 38% of variance in Gf. Further analyses suggested that only the latter testing conditions allowed low-capacity participants for relational learning during test taking, which allowed them to reduce their distance to high-capacity people. Study 2 corroborated the moderate value of WM–Gf correlation in untimed intelligence testing with a larger number of Gf and WM tasks, as well as showed that the indices of learning in a novel test of relation discovery predict significant amount of Gf variance. In sum, the research suggests that fluid reasoning can be differently related to WMC depending on the time pressure during Gf testing, and it also indicates that learning abstract relational representations may be an important component of unspeeded intelligence, but barely takes place during speeded testing.