What this is. 260 agents, 77 customers, three products, straight from the Standard Cost Credits 9/16 sheet. As it opens, every figure reproduces that sheet and the formula doc exactly; change any setting and the whole model recomputes.

Formula. Credits = ROUNDUP( cost / (1 - target margin) / price per credit ), clamped to the minimum. Standard cost prices a run’s own tokens at a fixed published schedule whatever model served it: Haiku 4.5 list, input $1.00, output $5.00, cache read $0.10, cache write $1.25 per 1M. Sonnet 4.6 is exactly 3x Haiku on every token type, so picking a schedule is the same lever as picking a margin.

The table. TJ’s columns in TJ’s order, under four headings. The orange column is the credit we charge; it follows the selector above, and every tile and chart is built from it. Actual cost minus Credits uses a minimum of 2, Standard Cost minus Credits a minimum of 1. Both margins are measured against what a run costs today, not against the standard. Diffs are new minus current.

Standard agents. Name matching is not enough. Each agent was read at its current version and compared word for word against 25 reference SOPs: the 16 published templates that carry a credit, plus 9 earlier standard SOPs still in use, so an agent on a previous standard still counts. An agent is tagged only if it keeps at least 80% of the standard SOP word for word and has not more than doubled it, the setting that best separates real copies from agents that merely share a name. 34 of the 260 qualify, against the 50 a name match claimed. Every agent’s score is in the table, so any row can be checked. Statement Reconciliation is counted at 7. The 8 Expense templates carry no credit, so Expense has nothing to check against.

Averages. Each agent counts once by default, which is how the sheet and the doc compute their tables. Weighted by runs weights each agent by how often it ran in the seven days to 16 September; 250 of the 260 ran that week.