Transparent sourcing
Open weights self-hosted, proprietary officially licensed
Every model states how it is supplied. There is no third option.
Whichever kind of supply you are buying, the page says which. "A bit of both" is not on the menu.
The real dividing line between relays is not price, it is where the capacity comes from. There are broadly four routes: buy the hardware and run open weights yourself, resell under a vendor’s official commercial licence, farm mass-registered consumer accounts into a pool, or reverse-engineer an official client protocol. The last two are cheap, and they are paid for with bulk bans that take your service down and with your conversations passing through layers nobody controls. We do only the first two, and every model page states which one applies — "on our own GPUs" or "official commercial channel". There is no third label and no model without one.
How it actually works
- Open weights are always self-hosted
- Officially released weights, on cards we bought, operated by us. Weight version and serving precision are published on the model page so you can re-test them with a pinned prompt set. We resell nobody else’s quota, so nobody else’s ban takes you down.
- Proprietary is always officially licensed
- Proprietary models cannot be self-hosted, so we go through the vendor’s official commercial channel: contract, invoice, audit trail, stated availability. That is a difference in kind from account pools and reverse-engineered protocols, not a difference in degree.
- Every model says which it is
- "How it is served" is a required field on every model page, visible at a glance. We would rather drop a model than list one whose origin we cannot state — that is exactly the thing we are telling you not to trust.
- What you buy is what you get, with nothing to wrangle
- One key reaches every model you have enabled, through one OpenAI-compatible endpoint, on one bill itemised per model. Switching models is the `model` field and nothing else — no second SDK, no extra credentials, no reconciling several invoices, no per-model retry logic.
Check it yourself
None of these four steps need us. You can run them right now.
- 1
Open any model page and look at the "how it is served" row. Every model has one; a blank is itself the signal.
- 2
Compare a proprietary model’s price against the vendor’s published rate. Anything far below it is not officially licensed supply — ask the provider directly whether it is a pool or a reversed endpoint.
- 3
Ask for a contract and an invoice. Account pools and reversed channels cannot produce either, which makes it the cheapest filter there is.
- 4
With one key, call an open model and a proprietary one in turn, check the `model` field on each response, and confirm the console itemises both.
The follow-up questions
Account pools are so much cheaper — why not use them?
Because you pay for the discount, not us: when the upstream accounts get banned in bulk your service simply stops, with no predictable recovery; your conversations pass through a layer you know nothing about; and the whole site can disappear one day. Closing that route off is what makes the word "stable" mean anything here.
What does "nothing to wrangle" actually mean?
It means you do not maintain different machinery for different models. One key covers everything you have enabled, one endpoint, one bill itemised per model. Changing models is the `model` field — SDK, retry logic and monitoring all stay put. Open and proprietary models are indistinguishable at the interface.
Could you quietly add a pool-sourced model one day?
Supply route is a public field on every model page, so adding a model means filling it in, and filling it in wrongly is a false statement you can hold us to. Any change is announced in the changelog first. And you can spot-check it yourself with points 2 and 3 above — which is precisely why we publish the field.
The other commitments
Don’t trust this page — test it
Take the trial credit and work through the verification steps above one by one. Anything that does not line up, send us the request ID.