Plug in your monthly call volume, token mix, and frontier price. kolm is a one-time compile cost, not a per-call cost. Year one absorbs the teacher bill once; year two and after, the same artifact runs at zero marginal token cost. The numbers below are unrounded and the formulas are the ones below the calculator.
Monthly call volume, average tokens per call, and your current frontier price. Defaults are the customer-support preset.
All numbers are dollars per year. Kept dollars compound; year-two onward the teacher cost goes to near-zero unless you choose to recompile.
In: per-call frontier cost (input + output token price), ~8% egress + variance buffer for the cloud lane, kolm plan cost, frontier teacher cost during the compile phase ((k - rows - in_tokens) priced at the same rate).
Out: engineer time, GPU rental during fine-tune (Pro tier handles this on the bridge), recall index population (one-time, scales with corpus size, not call volume), and the cost of not compiling (cloud variance during incidents, prompt regressions in production, vendor-side rate-limit incidents).
The honest assumption: tasks suited for kolm are tasks 80%+ deterministic on labeled examples. If your task is mostly free-form open-ended generation, the math swings; check /k-score for the gating logic, or open a thread in GitHub Discussions with your task shape before you trust this calculator.
A 30-day signed pilot. One workflow. One .kolm. K-score gate. The artifact stays with you whether or not we expand; that's the contract.
