How a blended rate is calculated
A rate card has several prices; a blend collapses them with weights. Choose what share of a million tokens is uncached input, cached input and output; multiply each share by its rate; add the results. The answer is what one million tokens of that traffic would cost on that model.
The weights are the whole story. Chat traffic sends short prompts and gets long answers, so output carries real weight. Agent traffic resends a large, mostly unchanged context on every step and writes comparatively little, so cached input dominates. The same model can therefore have very different blended rates for the two — which is why a blended number is only meaningful next to the mix it assumes.
The blend used on this site
Every effective rate on our model pages, comparisons and calculator comes from one formula in one place in our code, so the number cannot differ between pages. It models agent-shaped traffic: input is the large majority of tokens, output a small share, and most input is served from cache. The exact shares are printed in the panel below and change only with a dated methodology note.
One detail matters for fairness between vendors. A cache miss is priced at the model's cache-write rate when the vendor bills writes, and at the standard input rate when it does not, because a miss is the request that writes the prefix again. A model without a published cache-read rate is blended on input and output alone.
Artificial Analysis, which also publishes blended prices, assumes a 7 : 2 : 1 ratio of cache hits, uncached input and output — a similar, cache-heavy shape.
In our indexBlended rates in our index
Counted at build from our price index: the site-wide blended rate per million tokens among the 167 current public models we price, from 33 vendors. Each rate comes from the vendor's own price page, read between 2026-07-11 and 2026-10-09. How we collect prices.
What a blend cannot tell you
A blend is a ranking tool, not a quote. If your cache hit rate is low, your answers long, or your prompts cross a long-context threshold, your effective rate will differ from any published blend — sometimes by a multiple. Reasoning models complicate it further: hidden reasoning tokens bill as output, raising the output share in ways a fixed mix does not capture.
The practical use is in the middle of a decision: shortlist on the blend, then price your actual workload. The worked example shows how far the same model moves between two mixes.
Worked exampleSame model, two traffic mixes
10,000 requests a month on Claude Sonnet 5.5, priced at its current list rates, checked on 2026-10-09.
Read the blended line. The chat mix pays mostly output rates; the agent step pays mostly cache-read rates on far more tokens. Per request the agent step costs more, per million tokens it costs less — both numbers are true at once.
Change the workload or the model in the LLM API cost calculator.