WSJ: Only 11% of 396 Businesses Can Forecast Their AI Spend; Cheaper Models Costlier in 32% of Tests

The Wall Street Journal reports that just 11% of nearly 400 surveyed businesses can accurately project their AI costs as agentic workloads turn token consumption into a human-worker-style line item. A Stanford, CMU, Berkeley and Microsoft study running 6,800+ math, coding and science tasks across models found cheaper options actually cost more than higher-priced alternatives in 32% of scenarios — undercutting the assumption that cheaper-per-token models are cheaper overall.