I present a new inference-based analysis of belief attribution in which the embedded proposition is inferable from, but need not directly identify, an underlying belief of the subject’s. The analysis accounts for attributions of belief in necessary truths and falsities, overcoming a major difficulty facing Hintikka (1962), and goes beyond Cresswell and von Stechow (1982) in accounting for intuitively valid inference relations among belief attributions. The analysis is based on a novel I-semantics in which extensions are relative to the beliefs of a judge, according to an evaluator. The interpretation of believe uses syntactic inference, with premises deriving from beliefs of the judge (typically the attributor or evaluator) in addition to beliefs of the attributee. The proposed analysis makes no commitment to or use of possible worlds, and generates de dicto, de re, de qualitate and de translato interpretations depending on what gets raised out of the embedded clause.