In this experiment, we examined whether beginning readers benefit more from grapheme-phoneme decoding (GPD) than from whole-syllable decoding (WSD) instruction in learning to read and write words. Sixty Brazilian Portuguese-speaking first graders (M age = 6 years 1 month) who knew letter names but could not read or write words were randomly assigned to one of three conditions. The GPD group was taught to decode 40 consonant-vowel (CV) syllables by sounding out and blending grapheme-phoneme constituents (combinations of 10 consonants and five vowels). The WSD group was taught to decode the same CV syllables as whole graphosyllabic units. The individual grapheme-phoneme (IGP) group was taught the same 15 grapheme-phonemes as single units but no decoding. Groups were taught to a mastery criterion. Results showed that GPD instruction was much more effective than WSD and IGP instruction in enabling beginners to read CV syllables, multisyllabic words, and pseudowords; to learn to read words from memory; to write words; and to segment and blend spoken words phonemically. Also, GPD instruction facilitated phonological memory for spoken pseudowords. Despite receiving much practice in reading whole CV syllables, the WSD group learned few, if any, grapheme-phoneme subunits. Results support theories that reading instruction is most effective when it begins by teaching students to decode with small grapheme-phoneme units rather than with larger syllabic units, even when syllables are salient spoken and written units in the writing system.