Long-lived, deep-water species such as orange roughy (Hoplostethus atlanticus) are among the most difficult fish to age reliably, with experienced technicians processing only 60–80 otolith sections per day because of the complex, zone-based ring structure and severely compressed growth increments in older individuals. We present a multimodal deep learning framework that estimates orange roughy age directly from otolith section images, combining image features from a pretrained convolutional backbone with a Transformer-based encoder of the orange roughy ageing protocol and four biological covariates (fish length, otolith weight, section type, readability score). Three model architectures were evaluated; EfficientNet-B3, DINOv2-B and ConvNeXt-Large, were trained and compared on 270 annotated specimens spanning ages 20–117 years. ConvNeXt-Large, pretrained on ImageNet-22k, achieved the best performance in both image-only Mean Absolute Error (MAE = 3.13 yr) and multimodal configurations (MAE = 3.07 yr, RMSE = 4.05 yr; 64.4% of predictions within ±3 years), precision comparable with expert human readers. Prediction error was lowest for fish aged 35–40 years (MAE = 1.88 yr) and highest for fish older than 40 years, mirroring the difficulty experienced human readers face when resolving compressed post-transition-zone increments near the otolith margin. These results show that current machine learning models can approach the accuracy of expert human readers for one of the most challenging species to age.