The wrong unit
A prompt is not a stable unit of value. The same task can cost one model call or twenty controlled steps: find sources, flag conflicts, write a draft, check it and only then save. Charge per call and you gain financially from a high number of steps. Sell a cap on calls and you make customers afraid to use the system normally.
So Jarvis does not sell the underlying compute. The customer keeps their own relationships with ChatGPT, Claude, Codex or other providers. Jarvis sells the layer that decides which knowledge is relevant, which agent profile is allowed to work, where approval is needed and what the organisation remembers.
What does scale
The real product cost and complexity grow with the number of users, organisation brains, managed AI workers, version history, connectors, support and governance. Those are capacity limits you can understand. A team can work out in advance which plan fits without a calculator for tokens.
No free-and-unlimited promise
No TheSEO usage meter does not mean external providers are unlimited. A customer can have their own fair use, seat, model or usage terms with that provider. Jarvis does not make those costs an invisible part of one bundle and cannot guarantee the availability of a third party.
Why this improves how the product behaves
A fixed price per brain makes it sensible to keep source checking, approvals, evaluations and recovery steps rather than optimise them away. The interface does not need a credit meter counting down. Product analytics can look at time to an approved result, source quality, error reduction and reuse of knowledge you can rely on.
The check
The public pricing page and the machine readable `prices.json` should show exactly the same values. One change touches code, checkout, product copy, terms and the change register at the same time. Codex may flag the inconsistency and make a diff; a person approves the price and the effective date.