High impedance fault (HIF) is difficult to detect because of the low fault current amplitude and they can be easily confused with switching events. Existing HIF detection methods mainly rely on fixed time-window data. However, a fixed decision time often fails to balance the accuracy and speed of HIF detection. Thus, a sequence-adaptive HIF detection model is proposed in this paper. Firstly, zero-sequence current data of the faulty feeder are processed into variable-length training set. Then, a gated recurrent unit (GRU) model is trained based on variable-length data and cost-sensitive coefficient method to construct biased models. Two GRU models with opposite propensities are combined into an evaluation model. The test results on the PSCAD/EMTDC simulation platform show that the detection accuracy rate of this proposed model can reach 99.13%, and the detection speed is improved by at least 37.52% compared to the fixed time-window method. Delayed decision-making improves the accuracy of HIF detection and reduces the risk of harm.