Free until it works.
You fund a prepaid balance before your first request. Provider spend and our share are both debited from it, and every request comes back with a receipt showing each.
25% of what we can prove we saved you. Nothing if we prove nothing.
Charged on the request that saved the money, out of the prepaid balance you fund.
Explore
compare before you commitRun the same job across several providers, including ones you have no account with. You pay their prices plus a stated charge for running the comparison.
No share of savings. A comparison is never counted as a saving.
Managed Runtime
swap one keyYou pay provider prices out of prepaid credit. We add a share of what a cheaper model actually saved against the model you asked for.
25% of that saving. finest/auto is flat-priced and never share-priced, so nothing we choose for you can inflate a fee.
Enterprise
annual agreementSSO, replay inside your own VPC, audit exports and approval workflows.
A reduced, capped share of savings.
- We take 25% of the saving on the request that made it, debited as that request is served.
- Only a request we actually substituted can carry a share. Served as you asked it, served free of us.
- A swap that saves you nothing costs you nothing.
You were spending $100.00 a month. We recommended a better model that would have cost $150.00, and the routed mix we actually ran cost $90.00.
- What we bill against
- $10.00 the lower of your old spend and the option we recommended, minus what you actually spent
- Our fee (25%)
- $2.50 a quarter of that
- Your total
- $92.50 what you spent, plus our fee
- Extra value we do not bill for
- $60.00 what the pricier option would have cost above what you paid: shown, never billed as savings
Your bill went down, and it went down by more than our fee. Because we bill against the lower of your two numbers, saying yes to a recommendation can never raise the number we charge against. These figures are computed by the same function that bills, so this card cannot drift from the code.
The questions buyers ask first.
Does my traffic go through you?
Yes, and we will not pretend otherwise. You pointed your client at our gateway, so your request passes through our process in memory. We do not keep it: no table in this codebase has a column for prompt or completion content. Set FINEST_DISABLE=1 in your own environment and every call goes straight to your provider again, with no redeploy of our code.
What is the fastest thing I can do?
Change a base URL and a key, or tell your coding agent to add Finest. Nothing else. No route annotation, no test corpus, no graders. Every call is served as the model you asked for until a recorded test covers that exact job on a cheaper model.
How do you decide a cheaper model is safe for my task?
We start from tests Finest ran itself, on open corpora, and publish each result against the exact prompt, schema, endpoint and settings measured. Failures are published too, and they authorise nothing. A test we ran earns the label “Recommended”, never “verified for you”. Only a test on your own examples earns that.
What if the cheaper model is worse?
Then it is not eligible. A cheaper model has to clear a quality bar set before the test ran, on a frozen set of examples we name. Live traffic then starts with a small slice that rolls back automatically if results get worse. If the cheaper model refuses a call or returns malformed output, we re-run that call on your model at our expense.
Can accepting your recommendation make my bill go up?
No, and not as a promise. We bill against the lower of what you were spending and the option we recommended, so a pricier recommendation cannot become the number we charge against. It joins that number only if you independently run it in production for a defined window.
Is finest/auto a way to route me somewhere profitable for you?
Structurally, no. finest/auto is flat-priced and never savings-share priced, so the model it picks cannot increase what we charge. It rules out anything that fails your stated requirements first: region, modality, context window, output cap, streaming, structured output. An endpoint that cannot prove it meets a requirement is excluded rather than waved through. Every response names the provider and model that actually served you.
How do I leave?
Eject is a product, not a support ticket. It provisions your own provider accounts, exports your evidence and receipts, converts the policies we serve into policies you hold, and verifies your app no longer depends on our uptime. We advertise this because a business built on token margin cannot copy it without dismantling itself.
Do you need to see our prompts?
No. We test on frozen public, synthetic or CI examples only, and this release refuses uploaded, staging and production content outright. Encrypted content storage is not built, so we will not accept content we cannot protect.
Are you charging a savings fee today?
Yes, on requests where we actually substituted a model and the substitution saved you money. The share is 25% of that request’s saving, debited from your prepaid balance as the request is served, and it is on the receipt beside both prices. Two things never carry a share: a request served on the model you named, and a cheap attempt that failed and had to be re-run. And nothing is charged against a baseline you have not confirmed in the console.
Which models can you switch to?
Only models already in our catalogue, each with current published prices and a human sign-off. A vendor’s launch blog post is not a sign-off and cannot authorise a swap. We also need signed commercial terms with the provider before we will serve it for you at all, and that is off by default for every provider.