In Web 2.0, people are free to share their experiences, views, and opinions. One of the problems that arises in web 2.0 is the sentiment analysis of texts produced by users in outlets such as Twitter. One of main the tasks of sentiment analysis is subjectivity classification. Our aim is to classify the subjectivity of Tweets. To this end, we create subjectivity lexicons in which the words into objective and subjective words. To create these lexicons, we make use of three metaheuristic methods. We extract two meta-level features, which show the count of objective and subjective words in tweets according to the lexicons. We then classify the tweets based on these two features. Our method outperforms the baselines in terms of accuracy and f-measure. In the three metaheuristics, it is observed that genetic algorithm performs better than simulated annealing and asexual reproduction optimization, and it also outperforms all the baselines in terms of accuracy in two of the three assessed datasets. The created lexicons also give insight about the objectivity and subjectivity of words.