The complex pattern of eye movements during sentence reading has been computationally modeled using only three linguistic inputs, word length, frequency, and predictability. Whereas length can be easily computed and frequency can be obtained from databases, predictability is typically collected using the cloze task. Because this is labor-intensive and costly, cloze predictabilities are only available for a small set of texts. Recent developments in computational linguistics allow us to extract an alternative measure of predictability from open-weight large language models (LLMs), which are trained on next-token prediction. Thus predictabilities are now available for any text. Here we present evidence from lexical decision and eye movement experiments that LLM predictability strongly modulates ongoing cognitive word processing. Lexical decision times were substantially faster with high than with low predictability target words, and this was robust across variations in experimental procedure as well as in LLMs used to derive predictabilities. Given these findings, measures of word processing derived from eye movements should also be affected by LLM predictability. We evaluated this hypothesis using (a) a re-analysis of data from our longitudinal study of perceptual span development, for which predictability had previously not been available, and (b) designed experiments using the gaze-contingent boundary technique, in which we additionally tested whether predictability modulates semantic preview benefit. Results show that LLM predictability robustly influences gaze-based measures. Use of predictability increases with reading skill, has strong effects on early eye movement measures such as word skipping and first fixation duration, and increases interacts with semantic preview benefit