2026 phone class models reach 2024 frontier chat quality
2027 the split hardens; edge takes narrow high volume tasks, cloud keeps long horizon agents
2028 sub-4B on device matches 2026 frontier, on easy tasks only
Wildcard: a 100x compute efficiency breakthrough would collapse this gap. This is a projection, not a measurement.
A 2040 smartphone, will run something close to frontier quality for anything bounded and well specified. Frontier labs will plateau while the edge will keep closing. The capability index itself stops meaning anything. Every benchmark anchoring a 2024 = 100 scale is saturated well before 2040. Asking where it sits in 2040 would be like scoring a 2026 model on 2012 ImageNet. The split will survive anyway, relocated to compute budget rather than model quality. A chart that far out isn't worth generating, however as it is too large a timeframe to estimate for certain.