We address the challenge of dynamic option hedging using deep reinforcement learning (DRL). Unlike traditional model-based approaches, DRL is purely data-driven and does not require explicit modeling of the underlying market dynamics. Leveraging a comprehensive high-frequency dataset of SPX options, we train, validate, and test a DRL agent in a realistic market environment that incorporates actual transaction costs. The study highlights three key contributions. First, we analyze the hedging of 0DTE (zero days-to-expiration) options, which now dominate market trading volume, and show that DRL outperforms Black–Scholes delta hedging at this horizon. Second, we evaluate robustness across regimes, finding that the DRL hedge remains effective in crises such as COVID-19, even when trained only on non-crisis periods. Third, we examine the determinants of the performance gap, the role of alternative reward specifications, and hedging behavior in the presence of price jumps.