Tokens need a benchmark price

AI tokens are turning into a commodity. Commodities that trade well share one piece of infrastructure: a reference price that the sellers do not control. Tokens do not have one yet. This is the case for building it, and an honest account of how far STCI has got.

1. A commodity without a price

A million tokens from a capable model is becoming a unit of input that businesses buy the way they buy fuel or freight: in volume, from several suppliers, with a budget behind it. Guillermo Rauch put the moment as "more like an oil rush than a gold rush." Oil is a useful comparison because oil has something tokens lack.

Crude trades against Brent and WTI. Dry bulk shipping trades against the Baltic Dry Index, which the Baltic Exchange compiles each day from a panel of shipbrokers reporting what routes actually cost. No refiner and no shipowner sets those numbers. Supply contracts, budgets, loans and hedges are written against them.

Token prices work differently. Each provider sets its own price and changes it when it likes. A growing share of tokens reach buyers through intermediaries, gateways and resellers, some of whom subsidize the price to win volume. From the outside, nobody can tell a fair price from a subsidy. Nobody can budget a year ahead with confidence, and nobody can hedge, because there is nothing neutral to settle against.

2. What a benchmark needs

The working examples share four parts.

3. Grades: not every token is the same token

Brent is a grade of crude, not crude in general. A barrel of heavy sour is worth less than a barrel of light sweet, and the market prices the difference. Tokens have grades as well: the numerical precision the model is served at, the context it can use, the uptime and latency behind it. Two tokens from "the same model" are not necessarily the same product.

The neocloud spread is STCI's first measurement of grade. For each open-weight model it collects the price of every provider endpoint selling it and the precision each serves.

The spread measurement publishes with methodology 2.0.0. No spread file is in the published data yet.

A benchmark that ignored this would average fp4 tokens with full-precision tokens and call the result one price.

4. Where STCI is on that ladder

  1. Now Listed prices. Every separately priced model on one aggregator (OpenRouter), normalized to $ per million tokens, combined into four indices under a public rulebook. Each day’s value carries a verification hash that anyone holding the inputs can reproduce.
  2. Now Grades and spread. For open-weight models, the price of every provider endpoint selling the same model, and the precision each one serves.
  3. Next A reported-price panel. Gateways and large buyers report what they actually paid, anonymized and volume-weighted. Listed prices cannot show discounts, credits or subsidies; a panel can.
  4. Later Contracts that settle against the index. Forwards that let a buyer fix next year's token cost. This needs the panel and independent administration first.

Where it came from

I started STCI on January 1, 2026 and built the first version in a week: v0.1.0 on January 3 through v0.5.0 on January 8, with the collector, the indexer, the API and the site. It ran every day until March 6. Then I got pulled onto other work and switched the pipeline off, which is why the series has a gap. It came back on November 19, 2026 under methodology 2.0.0. The gap stays a gap: nothing in it is back-filled. Jeremy Longshore

What STCI is not, yet: it does not see prices paid, it uses a single source for listed prices, it weights models equally, and its 3:1 output-to-input blend is a stated assumption, not a measurement. Those limits are in the methodology, and the panel is the planned remedy for most of them.

5. The evidence

Today’s four index values and the full spread table are on the front page, and every number is in the JSON API.

6. Who it is for

Gateways and brokers, who see prices paid every day and would benefit most from a neutral reference to quote against. Enterprises that buy tokens at volume and want to know whether their price is fair. Finance teams that need to budget a year of inference. If that is you, report what you pay.