The lowest proven price,on every call you make.
One set of weights, a dozen price lists, and an invoice that rarely matches the work. Satei proves each host on the same questions, confirms the model is the one you asked for and the meter honest, then serves every request from the lowest price that held up.
Paying a host we do not list? Test it yourself, free.
94% below the mid price, listed by OpenInference at fp4. Mid is the middle of every price listed for this model, across the 30 hosts that carry it.
Prove
Each host faces the same questions until its identity and its meter are settled.
Rank
Only hosts that held up are ranked, and only by price.
Serve
Requests land on the lowest one, with nothing for you to watch.
Run the numbers
What your month would look like
Pick a model and your monthly spend. We price the identical tokens at the lowest listed host for it, fee and all.
Compared at today's listings, three parts input to one part output.
a month, about $11,298.41 a year
Today's widest gaps
Where the ledger spreads most
Per million tokens, three parts input to one part output. Mid is the middle of every price listed for that model, across all the hosts that carry it.
Model missing? It still flows through, marked unproven.
Open the ledgerThe loop
Prove, rank, serve. Nothing to check by hand.
One model is proven every six hours, taken in turn, so nothing on the ledger is older than a fortnight.
One question set, asked many ways
Every host answers the same six questions, twenty-four times over, in shifting wording. Real weights answer alike; a substitute breaks the pattern. Where we hold the tokenizer, each reply is counted against the meter too.
Prices ranked, never promised
Hosts that held up are ranked by price alone, and the winner goes on the ledger and into the free feed. No listing has to be taken on trust.
One base URL is the whole migration
Swap the base URL once. Requests land on the lowest proven host for your model, and move on their own the moment something cheaper clears.
The proof
Same weights. Wildly different invoices.
Identical model, identical answers, different host. Every row is a separate host listing the same weights.
Paying a host we do not list? Test it yourself, free.
Point the test at any OpenAI-compatible endpoint and watch its answers drift, in minutes.