Company

Why this exists

Finest exists because the model market reprices faster than any team can re-check it, and because everyone in a position to tell you so is paid by the traffic they would have to tell you to stop sending.

The market reprices faster than anyone can re-check it

A frontier model ships. A cheaper one matches it on your kind of work three weeks later. A provider changes how prompt caching is billed, and a long-context tier quietly reprices a whole class of request. Any one of those is an afternoon of work to evaluate properly. All of them, continuously, is a job nobody has.

So teams do the rational thing and pick a good model once. That decision is correct on the day it is made and slowly stops being correct, and nothing in the system ever tells them.

The obvious fix has an incentive problem

A gateway sits in the request path and can see everything needed to answer these questions. Every large one is free, because the traffic itself is the product: it feeds a spend platform, an adjacent cloud business, or a percentage. None of them can un-choose that position, and none of them makes more money by telling you to send less.

Finest is paid a share of what it can prove it saved you. If it recommends nothing worth doing, it is owed nothing. That is the only arrangement we could find where being right and being paid point the same way.

Evidence, or nothing

The hard part is not finding a cheaper model. That is a price table. The hard part is knowing whether the cheaper model still does your job, which nobody can answer from a leaderboard because your job is not on it.

So every claim in this product is attached to a record: a configuration, a frozen set of examples, a quality bar set in advance, and a result. If the record does not exist, the claim is not made. You will find that rule enforced all over this website, including where it costs us a nicer sentence.

We are in your request path, and we say so

Finest serves your traffic through our gateway. Your requests pass through our process. That is a real cost and it is stated as one everywhere it appears on this site.

What we do not do is keep them. No Finest table has a column for a prompt or a completion, and this release has no encrypted content store to put one in. One environment variable removes us, and everything we measured for you exports in a single file you keep.

Where this actually is

Finest is live and serving. Requests are debited from prepaid credit at provider list price, and a proven per-request saving carries our 25% share. The open-source packages are on npm and PyPI. The changelog is the honest version of the roadmap.

Two minutes of setup, then a receipt for every request.

Get a Finest key →See the argument in twelve questions →