( 01 — The brief )
Stop paying frontier prices for everyday work.
Run everything through the most powerful model and you pay top prices for tasks a cheaper one handles just as well. FrugalitiAI is the managed-credits app that fixes exactly that: type, and each request goes to the cheapest model that can do it competently, with the exact cost shown every time. The pricing is deliberately unglamorous — the real metered provider cost plus a small percentage, no subscription.
By the company’s own estimate, routing this way runs 40–75% cheaper than a top model for everything. It now also publishes a live counter of the measured saving across real accounts — around 68% in July 2026 — priced against what the same output would have cost on the top model, on a young user base that includes the makers’ own account.
— The client, in every case on these pages, was us. A demanding client — and, fittingly, a frugal one.
( 02 — The craft )
One window. The router decides.
It launched with two panes — chat and code — until a screenshot ended the argument: “how long will it take?”, typed into the build box, burned three agent steps before admitting it had no task. The panes were merged the same day. Now there is one box, and the server decides what a message deserves: an answer in words, a read-only look at the project, or a build.
The read-only mode earns its place: two labelling passes disagreed on every message like “is the search working”, so uncertain requests go where being wrong is free — nothing can be modified, a guarantee enforced where the action happens rather than requested in a prompt. Builds still run on your own machine, in a folder you choose, confirmed once per session.
( 03 — The engine )
Pay for the thinking only when it’s needed.
A message meets a cascade in which every step is gated on the last one admitting doubt: a cheap classifier types it, a cheap router picks the cheapest viable model and rates its own confidence, a stronger router is bought only when it isn’t sure, and a weak answer is redone one tier up — once, never an open-ended climb.
Every extra call — router, judge, re-answer — is metered into the single charge you see, so the cost of being careful is charged openly rather than buried. One rule disciplines the whole thing: the decision must cost far less than the work it routes. And where routing would be false economy it isn’t used — the build model is pinned, because a cheap one is prone to announcing it has finished when it hasn’t.
( 04 — Foundations )
It grades itself, and publishes the grade.
The classifier that reads your message is graded on hand-labelled cases before anyone meets it, and the harness never lets a single accuracy figure stand alone — the two ways of being wrong cost wildly different amounts. A wrong “answer” wastes one cheap reply; a wrong “build” turns an agent loose on your files. The score: 90.8% across 87 cases, with zero false builds across the 47 labelled “answer” — every miss on the safe side.
When a number didn’t survive scrutiny, the release notes said so in public — “I quoted a best run as if it were the value” — and the evaluation was rebuilt to report a range across repeated runs. After a release crashed on launch, twice, a check was added against the packaged installer itself, the thing people actually download. And the download page says it plainly: early beta, unsigned Windows installer, here are the two clicks past the warning — while the Mac button honestly reads “coming soon”.
Quote the range, not the best run.
( In brief )
- Free to try
- $1 credit, no card
- Choosing a model
- You never do
- Every reply
- Model, confidence, exact cost
- The work
- Runs on your machine