What's new
Analysis14 Nov 20265 min read

A cheaper tracker row may be a different seller

When yesterday's price came from one host and today's from another, a lower number does not prove either host cut its price, so the seller has to stay attached to every record.

Output priced at $2 per million tokens, then $1, is easy to summarise as minus 50%. The explanation needs another field. If the first record came from Host A and the second from Host B, all that was observed is a change in the quoted offer. It does not establish a price cut by the model's maker or either seller.

These are fictional prices and hosts, and we are not reporting a market trend. The point is to keep source continuity intact before writing a trend article, especially when a catalogue mixes direct lab prices with reseller listings.

Assumed US-dollar rates. Apparent change is −50%, but the seller changes between records. This is not a same-seller price history.
Fictional recordSellerOutput price per million tokens
Earlier recordHost A$2.00
Later recordHost B$1.00

What exactly is the price for?

A price record worth keeping names the model version, seller, endpoint, billing unit and service conditions. Input and output prices stay apart. A record for cached input, a batch route or another tier should keep that label instead of replacing the ordinary rate.

OpenRouter's provider-routing documentation, read on 14 November 2026, describes provider and endpoint selection and configurable fallbacks. A model identifier alone therefore does not identify a single seller. The route used for a comparison is worth saving, along with the resolved provider when your usage report supplies it.

In your own records, a change of seller or material conditions starts a new series unless you have a documented reason to compare it with the old one. Both offers can still be shown, as long as the label says what changed and a reader does not take the comparison for one company's pricing history.

Three dates that are easy to confuse

There is the day you fetched the source, the day the source says the rate took effect, and the day somebody last verified the interpretation. They can coincide, but they need not. Running the same import again shows only that it ran. It does not show that a hand-entered price was checked against the seller's page.

Where the source gives no effective date, the honest entry is unknown. A first observation tells you when you saw an offer, not the day it began. A missing row gets the same caution: it needs investigating before anyone calls it a retirement or withdrawal.

The source URL and the quoted unit should travel with the observation, because a bare number is hard to audit once it has been converted and rounded. Enough public evidence to reproduce the conversion is enough. Credentials, account details and private invoices do not belong in a public repository.

A bill estimate is not a price history

A bill estimate asks what a specified workload would cost at specified rates, and one dated offer is enough to work it out. A history asks how a comparable rate changed over time, which needs observations with a consistent meaning. A charting tool can draw a line between two points, but that does not justify the stronger claim.

If an application can switch sellers, the candidate routes can be compared on the same workload, with result quality, elapsed time and actual spend recorded next to the quoted rate. That tests two service configurations. It is not a statement that a lab made its API cheaper.

On a public chart, breaks in the series should show where the source changes, and missing intervals should look missing. A smooth line through an unverified gap implies evidence you do not hold.

When a price-cut claim holds up

We would want the same model version, seller, unit and relevant conditions on both sides, with dated source evidence. If those fields differ, the narrower observation is the one to describe: a different host is cheaper, a new tier exists, or the catalogue now uses a different source.

Done this way, a practitioner can see the offer they might use, and a reader following the history can see what the numbers support.