The Attic Standard Clause

The Attic Standard Clause ties the price of an inference commitment to an independent benchmark, so the rate follows the market rather than one vendor's price list. Buyer and seller agree the volume and a starting rate; each settlement period, the rate moves by the same proportion as the reference index.

Rate₀ is the contract rate at signing and Index₀ the index level on that date. Index_t is the level published for the settlement period, and a floor and cap agreed at signing keep the rate within a range both sides accept. The full model clause closes this page.

Inference prices do not drift, they jump. When a price changes, the median move is a third of the price, and a fixed-rate contract carries every one of those jumps, in either direction, for its whole term.

Experienced buyers know most prices hold from one week to the next; in the latest Attic Standard reading, only 1.8% of the 1,290 prices we track had changed. What matters is the size of the moves that do happen: the median change was 33%, and half of all changes fell between 16% and 52%. Three in four were cuts and one in four was an increase.

New models follow the same pattern. Almost one in five recently launched models was repriced after launch, and where the price fell, the median cut was 45%.

No buyer can know which vendor will move, or when. A contract signed the week before a cut locks in the old price for a year, and one signed before an increase leaves the seller carrying the loss until renewal.

The same model, on the same day, sells for up to 3.75 times as much depending on where it is bought. gpt-oss-120b is priced by thirteen vendors we track, from $0.00020 to $0.00075 per 1,000 output tokens.

This is not one odd case. Across every model sold by more than one company, the gap looks like this:

A vendor's list price tells a buyer what that vendor charges, not what the market charges. That is the gap an independent index fills: one published number, calculated the same way every week, that neither party to a contract sets.

A $10M annual commitment priced on Attic would have cost $48,000 less over its first five months than the same commitment at a fixed rate, and is now running about $317,000 a year below it.

The setup is a 12-month, $10M commitment to Flagship models, signed at the May base. The rate settles each month on that month's average level of the Attic Standard AIPI FLG GLB output index, with a floor at 90 and a cap at 110. Neither limit was reached.

The index fell 3.5% from its first May reading, and three single weeks carried most of that. Each month's invoice follows:A fixed contract bills $833,333 every month, $4,166,665 over the five months. Priced on Attic, each month's invoice follows the index:

Nobody reopened the contract to get there. Had Flagship moved the other way, as the Core index did over the same months, rising 1.2%, the seller's invoice would have risen with it, again within the cap and again without a negotiation. The figures scale in proportion to any commitment size.

The buyer gets market pricing for the whole term without renegotiating, and a cap that limits what a price increase can cost.

The seller trades a fixed rate for a longer, larger commitment, a floor under the price, and a share of any rise.

A settlement price only works if neither side sets it. Attic Standard sells no inference, buys none and routes none; its only product is the measurement.

The reference should match what the buyer actually consumes, which for most enterprises means a blend of published indexes weighted by their own spend.

One point to settle at signing: published indexes track the same models over time. They capture changes in price, not the savings from moving to a newer, cheaper model. A buyer who wants the latter can reference Attic Standard's spot measure in a tailored basket instead.

The language below is a starting point for the parties' own counsel to adapt, not legal advice.