The capability gap
Strong on the textbook. Weaker in the field.
94–98%
Frontier model accuracy on technical agronomy questions.
AI AgriBench v0.5, main track.
65–81%
The same models on real smallholder farmer questions.
AI AgriBench v0.5, Smallholders track.
The two tracks use different datasets and different sources. Score-to-score comparisons across them are not like-for-like. Read the key findings or how the benchmark is built.
Join the consortium
Organisations that can contribute relevant datasets, benchmarking, AI evaluation experience or traditional advisory expertise are encouraged to join.