The presentation duration of five-word sequences was varied and participants were asked to judge their grammaticality. The five-word sequences were presented for a variable duration randomly selected between 50 and 500 ms with 50-ms steps and were immediately followed by a masking stimulus. Half of the sequences were correct sentences which were randomly intermixed with ungrammatical sequences formed of the same words in scrambled order. We measured the proportion of correct responses for each presentation duration in the grammatical and ungrammatical conditions, and calculated sensitivity and bias from these measures. Both the sensitivity measure ( d′) and the probability correct responses to grammatical and ungrammatical sequences increased as the stimulus duration increased, with a d′ of 2 and an average percent correct close to 87% for the grammatical sequences already attained at 300 ms. The rate of increase in performance diminished beyond 300 ms. Grammatical decision times were faster and more accurate for the grammatically correct sequences, thus indicating that participants were not responding by detecting illegal word combinations in the ungrammatical sequences. On the basis of these findings, we provide an upper estimate of 300 ms as the time it takes to access reliable syntactic information from five-word sequences in French, and we discuss the implications of this constraint for models of reading.