275 Matched AI Listings Rose 4.4% on Input Price

Across 56 sealed dates, the matched OpenRouter input-price index moved from 100.00 to 104.44. The matched completion-price index moved from 100.00 to 105.04. Under this fixed-cohort method, the observed listing prices did not deflate. They rose.

That result is narrower than a claim about the cost of artificial intelligence. It describes one marketplace listing, one frozen set of model IDs, and two posted price fields. It does not adjust for model quality, capability, usage, latency, discounts, caching, or direct-vendor terms. Those limits are part of the result, not fine print added afterward.

The same audit also corrected a tempting but wrong headline. A raw scan found 169 model IDs whose full pricing objects changed across consecutive observations. After separating primary price fields from auxiliary fields, catalog movement, and gaps, the evidence supported 309 fixed-price events across 88 model IDs. One model can reprice more than once, so event count and model count answer different questions.

275 listing IDs form the fixed cohort across every sealed date.

309 adjacent-seal events changed a fixed prompt or completion price.

The compact JSON companion contains the full curve, method, diagnostics, CSV ledger hash, and four hashes for every named example below. The CSV companion exposes the curve and all ledger events as rows. Both downloads appear in the research-data panel on this page.

Why the original change count was not a reprice count

The pricing_json object contains more than prompt and completion prices. During this window, the source added or changed cache, image, audio, search, reasoning, override, and other auxiliary fields. A full-object comparison sees those changes. It cannot tell a buyer whether the ordinary input or output token price moved.

The audit therefore parsed every pricing object into stable economic fields before comparing dates. Key order was normalized. Decimal formatting was normalized. Catalog appearances and disappearances were kept outside the price ledger. A listing absent from either side of an interval became UNKNOWN for that interval rather than unchanged.

The 169 figure failed as the story because it mixed different kinds of change. It also included 1 changed observation reached only after a model was absent from an intervening seal. That gap is excluded from the adjacent-seal ledger. The source database also held 2 rows from an idempotent rerun that were not linked to the retained raw response for their date. Those rows are excluded rather than silently borrowed into the curve.

After those custody checks, 19,360 model rows remained linked to the raw responses. They span 449 distinct model IDs, but only the fixed, fully observed subset enters the index. This distinction prevents a sudden batch of new listings from moving the curve merely because the catalog composition changed.

The earlier USTA AI Price Index remains useful as a census of what was listed at an endpoint. This update answers a different question: how did posted prices move for the same continuously observed listing IDs?

The fixed-cohort method

The cohort contains a model ID only when that same ID appears on every successful source date with positive fixed values for both prompt and completion. Free listings, variable listings, invalid values, and any missing date keep an ID out of the cohort. The rule leaves 275 matched listings.

For each listing and date, the method divides the current price by that listing's price on June 11, 2026. It then takes the equally weighted geometric mean of those relatives and sets the first date to 100.00. Prompt and completion are calculated separately, but both use the same cohort.

Equal weighting is a deliberate limitation. The seal has no defensible usage, revenue, quality, or token-volume weights. A listing used heavily by customers therefore receives the same index weight as a listing with little usage. The result is a matched-listing price index, not a bill forecast.

The automation pricing-page audit uses an even stricter interpretation for static vendor pages. It treats changed dollar-figure sets as observations, not confirmed plan reprices. The two reports should stay separate because their sources and units differ.

The curve rose instead of falling

The matched input series ended at 104.44, a +4.44% change from its base. The matched completion series ended at 105.04, a +5.04% change. Calling this a deflation result would reverse the sign of the evidence.

The matched input index ended at 104.44.

The matched completion index ended at 105.04.

An increasing index does not mean every listing became more expensive. The event ledger contains 133 increase-only events, 148 decrease-only events, and 28 mixed events in which prompt and completion moved in different directions. Change frequency and index movement differ because the index also reflects the size of each relative move.

This is why the earlier two-census median could not answer the same question. A median over all paid listings changes when the catalog mix changes. The fixed cohort holds identity constant across the full 60-day span, so entry and exit cannot directly move the curve.

Dispersion is not a confidence interval

The public JSON reports the 10th, 50th, and 90th percentiles of listing-level price relatives on each date. On August 10, the input relative band ran from 100.00 at the 10th percentile through 100.00 at the median to 117.83 at the 90th percentile. The completion band ended at 100.00, 100.00, and 115.38.

Those ranges show dispersion within the complete qualified cohort. They are not confidence intervals. There is no probability sample here, so a sampling interval would suggest an uncertainty model the data does not have. Marketplace coverage, aliasing, equal weights, and missing capability adjustment are structural limits; resampling the same listings would not remove them.

The median staying at 100.00 while the geometric index rose is also informative. Most matched listings held their base price, while a smaller group of moves changed the geometric mean. The curve should therefore be read beside the event ledger, not as a substitute for it.

Three events and their custody trail

The examples below show why a ledger is more useful than a single aggregate. They include an increase, a decrease, and a mixed move. Prices are listed in USD per million tokens. Every example in the JSON includes its old and new model-row hashes plus its old and new raw-response content hashes.

The dates are adjacent successful source seals, not necessarily adjacent calendar days. That distinction matters for the qwen example because no successful source seal sits between its endpoints. If the listing had been absent from either endpoint, it would not appear as a reprice.

The hashes let an independent checker answer two questions. First, did each stored model projection preserve the listed economic fields? Second, did each row come from the retained raw OpenRouter response for that date? The builder aborts on either mismatch. It does not use a row merely because the row itself has a valid hash.

What announcement status can honestly say

The pricing clock archived daily model listings and several vendor pricing pages. It did not archive a frozen, exhaustive announcement corpus tied to each model event. A current web search would not prove that an announcement was absent when a change occurred.

For that reason, announcement status is UNKNOWN for the ledger, and the unannounced-event count is also UNKNOWN. This report does not label any event as disclosed or undisclosed. A future issue can answer that question only after an announcement universe, capture schedule, and matching rule are frozen before results are examined.

That boundary changes the headline but improves the product. A buyer can use the ledger as dated price evidence without inheriting a claim that the source material cannot support.

Put the Price Ledger to Work

A data or platform lead can use this ledger as an input to a controlled model-routing review. The useful automation is not “always choose the cheapest listing.” It is a change workflow that joins approved model IDs to dated price observations, flags a fixed-field move, and routes the event to the person responsible for quality, security, and commercial review.

US Tech Automations can build that workflow around an approved inventory rather than the whole marketplace. A practical record names the task, current model, fallback, prompt and completion profile, owner, evaluation evidence, and review date. The public ledger supplies a change signal; internal tests and contract terms determine whether anything should move in production.

For teams that also publish content, the AI-crawler blocking trend supports a separate access-policy review. Cost routing and crawler policy should have different owners and decision rules. Combining them into one score would hide what each source actually measures.

The research catalog maintained by US Tech Automations exposes the JSON and CSV without requiring a sales conversation. That makes the method inspectable before anyone asks for a tailored benchmark or monitoring workflow.

What this issue can and cannot support

This issue supports a dated statement about matched OpenRouter listing prices. It supports the event-level claim only when both adjacent rows and both raw response hashes reproduce. It also supports the finding that the broad 169-model change count was not a primary reprice count.

It cannot rank model quality, estimate workload cost, compare negotiated contracts, explain why a price moved, or generalize to the entire LLM market. It cannot treat an absent model as unchanged. It cannot infer announcement silence from a missing archive.

In this issue, nothing is estimated, modeled, annualized, or extrapolated. The geometric index is a deterministic summary of the frozen cohort, while the empirical percentile band is a description of that cohort rather than a sampling interval.

Frequently Asked Questions

Q: Did AI model prices deflate during this window?

A: Not under this fixed OpenRouter listing method. The matched input and completion indexes both ended above their 100.00 base.

Q: Why are there 309 events but only 88 repriced model IDs?

A: A listing may change on more than one adjacent-date transition. The event count measures dated transitions; the model count measures distinct listing identities.

Q: Why not keep the 169-model headline?

A: It counts any full pricing-object change across consecutive observations. Auxiliary fields and one nonadjacent observation are mixed into that number, so it is not a fixed prompt or completion reprice count.

Q: Is the percentile band a confidence interval?

A: No. It is the empirical spread of price relatives inside the complete 275-listing qualified cohort. No sampling confidence interval is reported.

Q: Does an event prove that a vendor announced or concealed a price change?

A: No. Announcement status is UNKNOWN because this issue has no frozen announcement corpus. The ledger proves the observed listing transition and its custody trail, not the vendor communication history.

Q: Can the index predict a customer invoice?

A: No. A real invoice depends on usage, routing, caching, discounts, direct contracts, and the models actually selected. The index is a public listing benchmark.

Source: US Tech Automations Research, sealed OpenRouter model-listing rows and raw-response hashes, June 11 – August 10, 2026.