An API business run by the people who use it.
PoolMill is a small, focused LLM API provider. We sell one thing: access to frontier models through a single OpenAI-compatible endpoint, at a lower effective cost than going direct.
Model quality converged. Model pricing didn't.
Most teams do not need a bespoke inference contract or a capacity reservation. They need one endpoint that answers reliably, bills honestly, and costs less per token than the alternative.
We buy inference capacity in advance at a discount and pass a share of that discount back to you as credit. That is the entire business model, and it is why a $5.00 payment buys $10.00 of usage rather than $5.00 of it. There is no spread quietly added to each token to make the arithmetic work.
Five commitments, each verifiable.
Published rates, metered per token
Input, cached input and output are counted separately and priced at the rates on our pricing page. Your console shows the request-level detail behind the total.
No dark patterns in billing
Credit is prepaid. When it runs out the API stops with a clear HTTP
429; we do not keep a card on file to charge, and nothing
auto-renews behind your back.
Prompt data is not retained
We store token counts, model, timestamp and cost so usage can be metered. We do not persist request or response bodies — verifiable in the database we run on, not just a line in a policy page.
Straight answers about limits
Upstream capacity is pooled, which is what makes the price possible. It also means sustained extreme burst rates can be rate-limited upstream at peak. We would rather say that plainly than let you discover it in production.
Your keys stay yours
Keys are scoped to the models you have paid for, can be created and revoked by you at any time, and stop working the moment you revoke them.
Kept small on purpose.
Requests arrive at an OpenAI-compatible endpoint, are authenticated and scoped to your account, and are routed to an upstream account pool. Token accounting happens per request and is priced on input, cached input and output independently — so the cache discount is a real line item, not an average.
Fewer moving parts means we can verify what we claim about metering, and a bug in our accounting is something we can find and fix rather than something you discover from your invoice.
Talk to us directly.
Questions about a plan, an invoice, or whether a specific model fits your workload — write to [email protected]. If something is broken, include the request ID: every response carries one and it maps directly onto the metering record.