Opetek Says Splitting AI Reasoning From Computation Cuts Query Costs 90%
Updated
Updated · InfoWorld · Oct 22
Opetek Says Splitting AI Reasoning From Computation Cuts Query Costs 90%
3 articles · Updated · InfoWorld · Oct 22
Summary
$0.07 per query versus $0.80 was the result in an Opetek benchmark when deterministic calculations were kept outside the LLM reasoning layer instead of feeding 10 years of EUR/USD data into context.
41% lower token prices from March to September 2026 have not solved the economics, because AI systems are consuming far more data as context windows expand from thousands of tokens to 1 million or more.
24-fold token demand growth by 2030, projected by Goldman Sachs, suggests bigger context windows alone will not keep costs in check if models repeatedly ingest raw datasets for tasks conventional software can compute cheaply.
Three efficiency layers now matter in production AI systems: model efficiency, retrieval efficiency, and computational architecture—deciding what the model never needs to read.
Capital markets highlight the problem early, but the same design tradeoff applies to genomics, physics, engineering, climate modeling and medicine as AI tools take on larger quantitative workloads.