Pricing and benchmark · 2026-09-08
Reading EU Model Prices: What to Watch Between EUR and USD
A qualitative guide to comparing prices when a European provider publishes in one currency and your catalog shows another, and to what actually drives an LLM bill.
The currency gap is the smallest part of the comparison
A European provider publishes its prices in its own currency while the catalog shows prices in one currency. That does not make comparison hard; the hard part is elsewhere.
The largest source of variance in an LLM bill is not the currency but the token distribution. The same number of requests on the same row produces a total that differs several-fold when the input-to-output ratio changes. Build the comparison on your own token profile, not on an exchange rate.
The line items that actually grow a bill
A bill has four line items, and unless you see them separately you cannot find where the money went.
- Output tokens: a multiple of the input price on most rows, and usually the largest part of the total.
- Reasoning tokens: produced whether you ask or not on some rows and billed as output; they escape notice because they do not appear in the answer.
- A server-side system instruction: some rows prepend an invisible instruction to every request, and it bills to you as input.
- Cache reads: repeated prefixes can be charged at a discounted rate, which lowers the total noticeably if you work with a long, stable context.
How to set up the comparison
Work from a real sample: take a hundred of your own requests, send them through two different rows, and record the input, output and cache counts from the usage that comes back. Multiply the price by those numbers; putting list prices side by side is misleading, because the same request produces a different number of tokens on each row.
Prices sit in the live catalog and on the model detail pages in their current form, which is why we print no figures here. A price table copied into a blog post becomes quietly wrong the moment the catalog changes.
Keeping the budget predictable
Three settings are enough for most teams: put a sensible output limit on every request, prefer rows that spend no reasoning tokens for batch work, and keep your stable context at the start of the request to raise cache hits.
Look at the first week's distribution in the usage panel. If the output line is larger than you expected, change the limit first and the row second — the order matters, because a limit change is free and its effect is immediate.
Frequently asked questions
Why is there no price table in this article?
Prices live in the catalog and can change. A table copied into a blog post becomes quietly wrong once the catalog is updated; check the model detail page for current figures.
How much does the exchange rate affect my bill?
Usually it stays small next to your token distribution. On the same workload, a shift in the output ratio moves the total far more than a currency difference does.
How do I benefit from the cache discount?
Keep the repeated context fixed at the start of the request. On rows that return cache hits the repeated prefix is charged at the discounted rate; the catalog shows which rows support it.