“The article is right to separate compute cost from retail price — but the retail price baseline itself is arbitrary depending on where you run the model. The same capability (e.g. Llama 3.3 70B with tool calling and 128K context) runs $3.00/1M tokens at model developer list price and $0.22/1M at Fireworks AI — a 93% gap for identical specs. That spread makes any “it costs Anthropic X” estimate depend entirely on which reference price you anchor to. We track this live across 1,625 SKUs and 40+ vendors at a7om.com — the variance across the market is larger than most people realise when they back-calculate provider economics.”