Where Data Tells the Story
© Voronoi 2026. All rights reserved.
Training a frontier AI model can cost anywhere from a few million dollars to several hundred million, depending on how much computing power a lab puts into it. The natural assumption is that more compute buys a more capable model, and a more capable model commands a higher price. The first half holds. The second half doesn't.
Group six frontier language models released between 2023 and 2025 by the raw computing power — measured in floating-point operations, or FLOP — that Epoch AI estimates went into training them, and three pairs emerge, each using roughly the same amount of compute. Inside every pair, list price still diverges sharply. Grok-3 and GPT-4.5 sit within 9% of each other on compute; Grok-3 lists at $15 per million output tokens, GPT-4.5 at $150 — ten times more. DeepSeek's V3 and R1 are within 6% on compute and still 2.4 times apart on price. The middle pair, Grok-2 and GPT-4, differ somewhat more in compute (41%), but the price gap between them — sixfold — still dwarfs it.
The same disconnect shows up between tiers, not just inside them: Grok-3 used 17 times the compute of the original GPT-4 and lists at a quarter of the price. What a lab spends building a model does not set what it charges for it — pricing strategy does.
Training-compute estimates are Epoch AI's. List prices are current OpenRouter floors for models still served, and archived launch pricing (Wayback Machine) for GPT-4.5, Grok-3, and Grok-2, all since delisted from OpenRouter.