How this is calculated

The median price across every host serving a model, not the headline rate.

The same weights sell for up to 3.6× different prices depending on who runs them, so a single published rate is a fact about one host rather than about the model.

Arena Elo from head-to-head build-offs, averaged across the categories a model was rated in.

Not an intelligence score, and not labelled as one. It is carried by 169 of the 432 priced models, 39% of them, so the chart and every value verdict on this page are drawn from that subset. The rest show a dash rather than a guess. The Artificial Analysis index is available as a second axis where present.

Nothing else is both cheaper and at least as good, inside a 18-point band on Elo.

Scores closer than the band are a tie, not a difference. The band scales with the axis: 18 Elo points against 2 on the Artificial Analysis index, since the two scales differ by an order of magnitude. Recomputed live from whatever workload and baseline you set.

Read per host, the context most providers actually serve, and the quantization each one runs.

Providers cap context below what the weights support and quietly serve compressed builds. A model offered as fp4 anywhere is flagged even when another host runs it at full precision.