Cardinality estimation and cost estimation can guide the selection of execution plan, and cardinality accuracy is very important for query optimizer. However, the cost and cardinality estimation techniques in traditional databases cannot provide accurate estimations because they do not consider the correlation across multiple tables. Recently, the application of artificial intelligence technology to databases (artificial intelligence for databases, AI4DB) has attracted wide attention, and the results show that the learning-based estimation method is superior to the traditional methods. However, the existing learning-based methods have some drawbacks. Firstly, most methods can only estimate the cardinality, but ignore the cost estimation. Secondly, these methods can only deal with some simple query statements, while do not work for complex queries, such as multi-table query and nested query. At the same time, it is also difficult for them to deal with string type of values. In order to solve the above problems, a novel method of estimating the cardinality and cost based on Tree-GRU (tree-gated recurrent unit), which can estimate the cardinality as well as the cost. In addition, an effective feature extraction and coding technology is applied, both query and execution plan are considered in feature extraction. These features are embedded into Tree-GRU. For the value of string type, the neural network is used to automatically extract the relationship between the substring and the whole string, embedding the string, so that the sparse string can be easily processed by the estimator. Extensive experiments were conducted on the JOB and Synthetic datasets and the results show that the performance of the proposed model outperforms the state-of-the-art algorithms in all aspects.