AI should cost less and know less about you

AI models have become a daily tool for writing, research, code, images, video and music. The best model for one job is rarely the best for the next, so the people who get the most from AI use several of them, often from several makers.

That habit has three costs that most people simply accept. Each maker wants its own account and its own bill. Each request carries more about the sender than the question needs. And most people pay the published price, because they are not large enough to negotiate another.

Moth is built on the view that none of these costs is necessary. The rest of this paper explains how we remove them, what that costs you, and what we can and cannot promise.

2.1 Many makers, many accounts

The models people rely on come from several makers. Using the right one for each job means an account with each, a key for each and a bill from each. Every account is another login, another card on file and another place where usage history builds up.

2.2 Identity rides along

A request usually travels with an account, an IP address and the details of a device, and many services keep the content as well. For a lot of work that is fine. For a contract, a health question or a draft nobody else should read, it is not.

The question has to reach the model. Who asked it doesn't.

2.3 List price for most buyers

Published prices are what individuals and small teams pay. The largest buyers negotiate their own, and smaller buyers rarely see the difference. The gap is not a secret, but there are few simple ways for a single developer to stand inside a larger pool.

3.1 One balance

You sign in by signing one message with a wallet on Robinhood Chain. There is no email and no form, and the server keeps only a keyed hash of the address. One balance, topped up with USDG or $MOTH, pays for every model in Moth Chat and through the API, with one Moth key.

3.2 Identity removed at the door

Every request reaches the model under a pooled key. Your wallet, IP address and client details are removed before it leaves our servers, so the host sees the question and not the person who asked it.

3.3 Nothing kept

We keep no copy of your prompts or outputs. Our own records hold only what billing needs, which is the model, the token count and the cost, for a period we publish and a deletion job enforces every hour.

A test in our code sends a marker string through every request path and fails if it shows up in any log or database table.

3.4 Priced against list

Pooled demand lets us price below the list price each maker publishes. We check every list price on the maker's own page and record the date, and we show the real saving beside each model, never a rounder number.

About half the models sit at half their list price. Others sit closer to it, and a few have no list price to compare against, so no saving is claimed for them.

Savings are easier to judge as a month of real use than as a price per million tokens. The figure below prices five sample months from our catalog, each against the same use at list price.

The samples are illustrations, not forecasts. The calculator on the pricing page prices any model at any level of use.

Every price in this chapter is read from our catalog when the page is built. Price your own month.

Every model carries one of two labels, and each says exactly what it hides.

5.1 Anonymized

Anonymized models hide who you are. Your wallet, IP address and client details are removed, and the frontier models run with their makers or partners, who can read the content under their own policy.

5.2 Sealed

Sealed models hide who you are and what you send. A few open-weight models run in attested hardware, the content is encrypted to it, and neither the host nor Moth can read it. Sealed arrives at launch, and each request passes four steps in order.

5.3 What we can't promise

• The hosts' own policies. How the hosts of Anonymized models handle content is governed by their policies, not ours.
• What a prompt says. Removing your identity doesn't make the text anonymous. A prompt can still identify you by what it says.
• Payments on a public chain. Payments on a public chain are visible, so we say no account is needed rather than calling it anonymous.

5.4 No filter added on top

We add no moderation or prompt filter of our own. Frontier models behave as their makers ship them, and open-weight models with lighter defaults sit beside them. The terms still ban illegal use, and every key has rate and spend limits.

We say all of this plainly because a privacy claim that turns out false costs more than a smaller claim that holds.

Every call holds its worst case before the model runs, so a balance never goes below zero. When the reply ends, the call is charged for what it really used, and the rest of the hold goes straight back. A call that fails releases all of it.

Every reply carries its request id and exactly what the call cost, in micro-dollars, in a header named moth-charged-micros. Prices never change because of routing, and if a model can't be served at its price, the request fails with a clear error.

$MOTH ties the people who hold it to the use of the platform. Every swap pays a fixed 3% tax, and by default a quarter of it goes to holders each hour, valued in USDG, to claim as credit for models or withdraw as USDG.

The split goes through a splitter contract. At launch it burns 0.25%, sends 0.75% to Holder Credit and 2% to development and operations. The owner can change the split, and every change is recorded on-chain where anyone can read it.

This is not an offer to sell, or a solicitation to buy, any token or security, and nothing here promises a price, return or profit. Claimed credit is prepaid access to the platform.

• No model gets smarter. None of this needs a new model. The value is in how the system around the models is built.
• No speed or accuracy figures. We publish none until our own reproducible test backs them.
• No price promises for the token. It is described for information only.
• No anonymity we can't give. We promise only what our own servers do, and say where that stops.

The thesis holds if developers move real workloads to one Moth key and keep them there, and if people bring Moth Chat the questions they wouldn't type anywhere else. We will publish what we learn, including where it doesn't hold.

Until launch, every number in this paper comes from our catalog and our own published rules, and the sources below are the pages each list price was checked against.