Neural networks particularly long short-term memory (LSTM) neural networks have found a way into time series prediction models and are seldom applied in financial option pricing. Unfortunately, there has been a lack of sufficient research on the pricing capabilities and efficacy of LSTM neural networks in the realm of financial options. This research analyzes the S&P 500 index options data employing the following models: standard LSTM, bidirectional LSTM, stacked LSTM, GRU, and bidirectional GRU. Our approach in constructing the input features includes option Greeks, rolling features, shifting features, and all other features computed from the price data. Feature importance is ascertained through random forest analysis and used to choose ten features for training the model. This approach ensures a focused and relevant input set for each LSTM variant. The performance comparison shows that the proposed GRU model yields higher accuracy than other models in two assessment datasets including call and put options. To undertake these tests, statistical hypotheses are developed that make it possible to determine if the differences observed in the models’ performance are statistically significant. Empirical results demonstrate that the GRU model consistently outperforms other variants across both call and put options datasets. To validate these findings, we conduct rigorous statistical tests to assess the significance of performance differences among the models.