One flat fee per task. No per-token pricing, no usage meter, no subscription required, a local model that charged you per request would defeat the point.
Everything it takes to get one task from "we do this by hand / via a cloud API" to a small model running inside your network, done by a dedicated team of two engineers.
You get one number in the proposal. It doesn't change unless the scope does.
Drift monitoring, periodic retraining, base-model upgrades. Also flat, monthly, cancel any time, the model keeps working without us.
USD, excluding taxes. Russian-market engagements can be contracted via a local entity in rubles. Hardware is not included, most clients run it on a server they already have.
A back-of-envelope the sales team will happily do with you on a call.
A document-processing workload of ~2,000 tokens per item at 300,000 items/month is ~600M tokens/month. At typical frontier input pricing plus output, that is on the order of $1,500-15,000 every month depending on model tier, and the price is not yours to set.
Same workload on one server you own: electricity and amortized hardware, typically a few hundred dollars a month, flat, no matter how much volume grows. Payback on a production build is usually measured in months, and the compliance benefit arrives on day one.
Tell us the task in a sentence. We'll reply within one business day with whether it's a fit, and a proposed metric.
Or email 4rd@atknin.com