AI evaluation of science images… highest 64%, cost gap 99 times

Google's Gemini 4 Carbon topped Artificial Analysis's independent VIALS-AA benchmark of 19 AI models on life-science image interpretation, scoring 64%, while the evaluator reported a 99-fold gap in per-question cost between the cheapest and priciest models. The test, an independent implementation of Handshake AI's VIALS standard, drew on 161 visual question-answering tasks spanning gel blots, plasmid maps, microscopy images, flow cytometry plots, phylogenetic trees and molecular structures. The results matter because accuracy on lab imagery varies widely across models while inference costs diverge by two orders of magnitude, a key trade-off for research teams deploying these tools. What to watch next is whether the top-scoring models can narrow the cost gap and how the benchmark evolves as more models are added.