This paper investigates the potential of the vision-language pretrained model contrastive language-image pre-training（CLIP） in few-shot open-set recognition （FSOR）. The experiments reveal that the visual-prototype classifier based on CLIP’s image encoding features generally performs worse than traditional FSOR baseline methods； although the semantic-prototype classifier based on CLIP’s semantic encoding features significantly outperforms traditional baselines in closed-set performance， it underperforms in open-set performance. The primary reasons for these issues may be the gap between CLIP’s training data and the FSOR target data， as well as the overly large decision boundaries assigned by the CLIP semantic prototype classifier to known classes. To tackle these problems， a simple yet effective joint few-shot open-set classifier enhanced with visual-semantic integration is proposed， which not only leverages the closed-set classification advantages of the semantic-prototype classifier based on CLIP but also skillfully exploits the potential of the visual-prototype classifier constructed by traditional FSOR pretrained models， which further enhances the open-set performance by establishing tighter decision boundaries. Experiments on four benchmark datasets demonstrate that this method achieves average improvements of 2.9% in accuracy （ACC） and 2.6% in area under the receiver operating characteristic （AUROC） compared to the best traditional baselines.