The Spatial-Numerical Association of Response Code (SNARC) effect refers to the finding that small or large numbers elicit faster leftward or rightward responses, respectively. Traditionally, this effect has been thought to reflect the intrinsic spatial orientation of the mental number line (MNL account) and thus to be modulated by the amount of semantic processing required by the task. This study aimed to test this hypothesis. Participants performed two tasks requiring semantic processing (magnitude classification and parity judgement) and two tasks requiring the processing of non-semantic features of the number (phoneme detection and color judgement). Contrary to the MNL account, the SNARC effect in the four tasks was not modulated by the amount of semantic processing, but rather by response latency. These results provide evidence against the MNL account and in favor of alternative accounts (dual-route model, working memory account).