Inference and sandboxes for autonomous agents. Metered per unit, capped per key — and the cap is taken before the call runs, so a runaway agent is refused rather than reconciled.
Every call holds its worst possible cost before it runs. Two calls cannot both believe there is room, because the room is taken rather than read.
The hold settles afterwards to what was actually used. Settlement can only give money back — an upstream that over-reports is clamped to the hold.
A call that would breach the cap returns a refusal naming the shortfall. Nothing ran, nothing is owed, and there is no invoice to dispute later.
A token can cost a millionth of a cent, so cents are not small enough to hold the price and floats are not exact enough to hold the total. Everything here is an integer number of micro-cents, start to finish.
| Model | In µ¢/tok | Out µ¢/tok | µ¢/sec |
|---|---|---|---|
| swork-fast | 2 | 8 | 0 |
| swork-deep | 6 | 30 | 0 |
| swork-sandbox | 0 | 0 | 900 |
swork-fast, 4,000 in, ceiling 2,000 out
4,000 × 2 = 8,000 µ¢
2,000 × 8 = 16,000 µ¢
────────────
reserved 24,000 µ¢ = $0.2400⚠ The output figure is the ceiling you declared, not an estimate. Reserving what a call is likely to cost and hoping is the same bug as checking the balance — the call that overruns is precisely the one you needed to stop.