AI Models Identify Nearly Half of 2,000 Spider Photos at Species Level

Summary

A benchmark tested nine AI models on the same 2,000 spider photos covering 671 species and subspecies. Each model selected the species from a list of 20 possible names, using research-grade iNaturalist observations and community identifications as the expected answers. Gemini 3.8 Flash ranked first with 997 correct answers, or 49.85%, followed by GPT-6 Astra at 47.50% and Claude Fable 5.1 at 43.10%. The results were specific to this dataset and were not presented as a universal model ranking. The errors also showed useful partial recognition: Gemini placed 871 of its 1,003 misses in the correct spider family, while GPT-6 Astra and Claude Fable placed 92.00% and 91.45% of all answers in the correct family. Muse Spark 1.3 ranked fourth but was far cheaper, costing an estimated $0.93 for the full run versus $27.11 for Gemini, whose 215 additional correct answers cost about $26.18 more. The author notes that exact species identification can require details absent from ordinary photos, so family recognition and species-level accuracy measure different capabilities.