We present a computational model of sentence production thatemulates variation of the output of lexicalization andgrammatical encoding of the abstract pre-lexical message, interms of complexity and accuracy of the generated sentence aswell as fluency and cognitive costs of the sentence production.The model integrates approaches from routine action selectionmodels built on Dual Systems Theory (Norman & Shallice,1986) with ‘A Blueprint for the Speaker’ developed by Levelt(1989). The paper describes and justifies the modelarchitecture, explores factors affecting language variation inproduction, and applies the model for testing relationshipbetween complexity, accuracy, and fluency (CAF) of languageproduction as debated within Second Language Acquisition(SLA) research. A simulation that generated 78,750 sentencesprovides evidence of the trade-off relationship between CAFparameters as speakers have to sacrifice performance on one ofthe CAF factors in order to improve the remaining two.