TokenCostEstimate bill →
Market Index

The LLM API Token Price Index

A single number tracking how blended LLM API pricing has actually moved since Jul 2026 — reconstructed entirely from our own real, dated price history, not estimated or extrapolated backward.

109.6+9.6% since Jul 2026base = 100 · 6 weeks of real data · 108 models tracked

How this index is built

Every point on this chart is reconstructed from our own price-change log — not a separate estimate. We start from today's real, live pricing and walk each dated change backward, so every historical point reflects prices we actually had on that date, not a model or a guess. Three design choices, disclosed rather than buried:

  • 3:1 input:output blend. Each model's price is blended as 0.75 × input + 0.25 × output per 1M tokens, reflecting a realistic mixed-usage pattern rather than input-only or output-only pricing. This matches the convention used by other public LLM price indices, so figures are comparable across sources.
  • Median, not mean. The index is the median blended price across all tracked models in a tier (or overall), so one outlier launch or one aggressive discount doesn't swing the whole number — you're seeing the market's center, not its extremes (those are covered on the trends page itself).
  • Rebased to 100 on Jul 2026 — our actual first tracked date. We deliberately did not extrapolate this backward to an earlier "base year" the way some indices do; we only show index values for dates we have real data for.

What moves this index that isn't a "price change"

Because the index tracks the tier's median composition, adding or losing a model changes the index even when no existing model's price moved — a new expensive flagship entering the Premium tier pulls that tier's median up, independent of any provider cutting or raising a price. This is a known, disclosed limitation of every price index like this one, not unique to ours; read a sharp single-period move alongside the event log before concluding prices themselves shifted.

Why the history is short — and why that's intentional

This index only goes back to Jul 2026, 6 weeks of real data as of today. That's genuinely a short window, and a single downtick or uptick over a few weeks isn't a trend — it's one data point. We chose not to backfill a longer series with estimates, because the whole point of this index is that every number in it is real. The series will get more useful as it gets longer; check the "updated" date above to see how much history exists right now.

Frequently asked questions

Why 3:1 input to output, not something else?
It's a rough approximation of typical mixed usage, and it matches the blend ratio used by other public LLM price indices, which keeps our numbers comparable to theirs rather than using an idiosyncratic ratio nobody else does.
Does a rising index mean prices are going up?
Usually, but not always — the index can also rise because a new, pricier model entered the tracked set, shifting the tier's median composition. Check the price-trends log for the same period to see whether it was driven by actual repricing or by which models were active.
Why doesn't the index go back further than mid-2026?
Because that's when our daily pricing pipeline actually started recording dated snapshots. We could have extrapolated a longer series from press releases or memory, but that would mean some of the "real data" claim on this site stops being true — we'd rather show a short, fully honest series than a long, partly invented one.
Are retired models included in the history?
Only if they're still in our tracked lineup today — the index is reconstructed by walking today's real prices backward through dated changes, so a model retired before today isn't part of that walk-back. This is a real limitation we're disclosing rather than working around with an estimate.