Recommender systems are essential to various fields, e.g., e-commerce, e-learning, and streaming media. At present, session-based recommendation models are normally based on the next item recommendation, which recommend items existing in users’ historical sessions. However, recommending items that users have interacted (old items) will reduce the interest of users interacting with new items and leads to an information cocoon. Therefore, it is necessary to recommend items that users have never interacted (new items), namely, session-based new item recommendation problem. This problem is new and challenging since new items have no interaction with users. It is difficult to learn the representation of new items and recommend them to users effectively. Motivated by these challenges, we propose a dual-intent enhanced graph neural network for the session-based new item recommendation. User intent expresses the user preferences of items, which is the core part of the recommender system. To learn the user’s intent effectively and determine the range of new items that are likely to be recommended to users, we use the user’s historical session to learn the user intent by an attention mechanism and the distribution of historical data, respectively. Since new items have not interacted with the user, inspired by zero-shot learning (ZSL), we infer the new item representation by using their attributes. Finally, by outputting new item probabilities, which contain recommendation scores of the corresponding items, the new items with higher scores are recommended to users. Experiments on two representative real-world datasets show the superiority of our proposed method. The case study from the real-world verifies interpretability benefits brought by the dual-intents and new item reasoning.