Craftax is a new benchmark for open-ended reinforcement learning (RL) environments, designed to address the limitations of existing benchmarks. Existing benchmarks either lack computational efficiency or are too simple to pose significant challenges. Craftax-Classic, a JAX-based rewrite of Crafter, runs up to 250 times faster than the original Python implementation, allowing for more extensive exploration with limited computational resources. The main benchmark, Craftax, extends Crafter with elements inspired by NetHack, adding complexity and depth to the environment. Solving Craftax requires deep exploration, long-term planning, memory, and continual adaptation to novel situations. Experiments show that existing methods, including global and episodic exploration, and unsupervised environment design, make little progress on Craftax. The authors believe that Craftax can serve as a meaningful challenge for future RL research, enabling experimentation in a complex, open-ended environment with limited computational resources.Craftax is a new benchmark for open-ended reinforcement learning (RL) environments, designed to address the limitations of existing benchmarks. Existing benchmarks either lack computational efficiency or are too simple to pose significant challenges. Craftax-Classic, a JAX-based rewrite of Crafter, runs up to 250 times faster than the original Python implementation, allowing for more extensive exploration with limited computational resources. The main benchmark, Craftax, extends Crafter with elements inspired by NetHack, adding complexity and depth to the environment. Solving Craftax requires deep exploration, long-term planning, memory, and continual adaptation to novel situations. Experiments show that existing methods, including global and episodic exploration, and unsupervised environment design, make little progress on Craftax. The authors believe that Craftax can serve as a meaningful challenge for future RL research, enabling experimentation in a complex, open-ended environment with limited computational resources.