Put in more at once, and we add more on top.

Because the two licensed modes are quoted against an estate (how many models, how many environments, how much of it we operate), and a number printed here would be wrong for almost everyone reading it. Cloud is metered on tokens at published rates, which you can work out for your own volume in the calculator on the home page.

One credit is one dollar of routed inference, priced per request against a versioned price table. The version that priced a call is recorded with it, so a bill can be reconstructed months later against the rates that were actually in force.

No. Credit is drawn down as you route and sits there when you do not. Nothing resets, and nothing is forfeited at the end of a period.

The first top-up is $1, charged to verify a card, and we add $4 of bonus credit on top of it: $5 to start. It routes against the full estate at the ordinary rates (it is not a reduced tier or a sandbox), so what you see from it is what production will do.

No. It routes at the same rates, draws down in the same order, and does not expire. The bonus attaches to the top-up rather than to your account, so it is earned per purchase and there is no tier to maintain.

No. A failed attempt costs you nothing, and the fallback that replaces it is never billed above what the original request reserved. If routing costs more than the model you asked for would have, we absorb the difference.

No. On premises and hybrid licences are priced on seats or routed volume as agreed in the quote; what your own hardware spends serving a request is yours, and the per request accounting prices it against a table you supply.

The price table is versioned and dated, and announced changes are scheduled rather than applied by hand. A rate with a published start date takes effect on that date on its own. Calls priced before it keep the version that priced them.

Through Elytra: watched on every release, probed on our own hardware and scored by lower bound, cleared for licence and deployability, then signed off by a reviewer. Around 750 under watch resolve to an approved estate of roughly 460.

Yes, in all three modes. A private endpoint or a self hosted deployment passes the same gate as everything else in the estate, and routes on the same evidence.

Capability first, then cost. Context, output length, and required skills eliminate the models that cannot do the work; among those that can, the router takes the cheapest that clears the quality bar, and raises that bar for code and deeper reasoning.

Yes. Name one and it is used, with routing reduced to failover within that model's own providers. Policy can also fix the choice for a class of work while leaving the rest to route.