Estimate monthly inference spend across four common enterprise functions, then see how frontier labs compare to sovereign and self-hosted open weights, energy included. Refreshed August 2026 on published prices and the Artificial Analysis Intelligence Index. A directional model, useful in five minutes.
For each function, we show monthly cost per model and a directional performance score. The right answer is rarely the cheapest, it's the cheapest model that clears the quality bar for that task.
Sources, August 2026: frontier pricing from Anthropic's published API rates; open-weight pricing and the quality index from Artificial Analysis. The self-hosted line is modelled rather than quoted: the same Qwen3.8 27B weights rent at $0.40 in and $3.00 out per million tokens on a public host, and owning the hardware removes that margin, which is the whole comparison. The open-weight table moves every few weeks, so treat any figure here as a starting point rather than a quote.
We'll send you a PDF with your scenario, refined with realistic GPU capex, energy under EU industrial tariffs, ops headcount, and the break-even point against frontier-only spend. Plus the per-function recommendation we'd give in a working session.