What the labs charge

These figures are not directly comparable. Each is a price observed on a particular day, and those days span 847 days — from 2024-06-20 to 2026-10-15. Two figures read years apart are not a price comparison, and published prices are cut without notice, so any of these may be stale. The observation date is in every row for that reason, and there is no chart ranking these against each other, because a bar chart would assert a comparison the data cannot support.

Published prices, per million tokens

Newest observation first. USD, per million tokens, as published by the lab. Where a record carries several observations, the most recent is shown.

Price against when it was read

The same figures plotted against their observation date, log scale. This is the confound made visible — but not in the direction you might expect. Fourteen of the sixteen were read within one recent stretch, and the vertical spread inside that stretch is far wider than the gap to the two older readings: prices within a single week range from $1.25 to $75, because they are different tiers of model, not different eras. The date matters for whether two figures can be compared at all; it does not, on this data, explain the spread.

What is missing

236 of 280 records carry no pricing at all. Open-weights releases often have no list price to record, and several proprietary labs publish pricing only on pages that change without notice — which is why every figure here is cited to a dated snapshot rather than a live page. No record carries more than one observation yet, so price cuts are not yet visible; the schema holds a list precisely so they can be, once a second reading exists.