Oxytocin (OT) has been implicated in the formation and maintenance of various social relationships, including human romantic relationships. Competing models predict, alternatively, positive or negative associations between naturally-occurring OT levels and romantic relationship quality. Empirical tests of these models have been equivocal. We propose a novel hypothesis (‘Identify and Invest’) that frames OT as an allocator of psychological investment toward valued, vulnerable relationships, and test this proposal in two studies. In one sample of 75 couples, and a second sample of 148 romantically involved individuals, we assess facets of relationships predicting changes in OT across a thought-writing task regarding one’s partner. In both studies, participants’ OT change across the task corresponded positively with multiple dimensions of high relationship involvement. However, increases in participants’ OT also corresponded to their partners reporting lower relationship involvement. OT increases, then, reflected discrepancies between assessments of self and partner relationship involvement. These findings are robust in a combined analysis of both studies, and do not significantly differ between samples. Collectively, our findings support the ‘Identify and Invest’ hypothesis in romantic couples, and we argue for its relevance across other types of social bonds.