Inference price indices built from the published reference prices. Pick a family, a measure and a segment in the bar below; it stays at the top as you scroll. Every index is chain-linked points with a USD basket cost: methodology v0.3 for core, equal-model and unit value; v0.4 for settlement grade, built to be referenced by a derivative rather than just observed. The equal-model tiers rank capability by model count instead of by tokens, to show how much the tier definition matters. All are experimental. For why the cuts are drawn this way and whether they hold up, see Analysis.
Custom blend
Your own input/output ratio, for the segment selected above.
Answers None of the five standard blends match my workload — what would I actually pay today, and is that getting cheaper or more expensive?
50% / 50%
—
USD per 1M tokens, today
—
vs the 50/50 blend
Computed live from the segment's published input and output basket cost — not a stored series, and not chain-linked, so there is no reconstitution math to get wrong: it is exactly alpha × input cost + (1 − alpha) × output cost on the same two numbers behind the 80/20, 50/50 and 25/75 measures above.
Change in each index
How far each index moved over the period. The indices are in points and the unit value in dollars, so they compare as percentages only; the unit value also moves when usage shifts between models. Click a bar to open that series.
Answers Which segments moved the most over this period, at a glance across all of them at once.
Settlement grade v0.4
Built to be referenced by a derivative — a future, a swap, a term structure — not just observed: weekly reconstitution, an incumbency buffer on selection, a disclosed exception for a member that goes dark between reconstitutions, and Premium/Standard/Economy tiers ranked by price rather than capability. See Methodology.
Core, equal-model and unit value
Every series in these three families at a glance, with its latest level and 1-fixing change — settlement grade's own board is above. Click any cell to open it.
Core indices
Index points (base 100), 1-day change, and today’s basket cost in USD per 1M tokens. Tiers rank AA Intelligence by consumed tokens.
Equal-model tiers
The same construction, with AA Intelligence ranked by model count: each release is one step, however much it is used. A sensitivity check on the core tiers.
Market unit value
The day’s quoted spend divided by its tokens: each provider’s price times the tokens it served, over all published releases. A token-weighted average quoted price (each token counts once, whatever it cost), the construction of Silicon Data’s SDLLMTK. It moves with the usage mix, so it is not a price index.