The Orchestrator

The percentages above need a bill. Here is one, itemized, for a delegated topology of the kind Chapter 7 defined and Maya’s team in Chapter 2 would find familiar: four agents doing the ongoing work of a mid-sized internal operation — triage and classification, retrieval and drafting, review support, and a heartbeat monitor — plus one human-reviewed decision step. The arithmetic uses list rates without caching or batch discounts and excludes hosting and separately billed tools. Output volumes include any billed reasoning tokens.

The workload: 12,000 classification calls, 8,000 retrieval-and-draft cycles, 4,000 review-support summaries, and 30,000 heartbeat checks a month. The cheap model is DeepSeek V4.1 Flash off-peak, at $0.15 per million input tokens and $0.60 per million output; the frontier model is Claude Fable 5.1, at $10 and $50. Everything runs cheap except the two steps where judgment pays for itself: 5 percent of the drafting tokens and half of the review support go to the frontier model. Assume malformed outputs add 4 percent more attempts, with the same average token volumes and model mix.

The arithmetic, in one pass. Classification is 12,000 calls averaging 1,500 input and 300 output tokens, cheap only: 18 million tokens in and 3.6 million out, about $5. Retrieval and drafting is 8,000 cycles averaging 6,000 input and 2,000 output tokens, 48 million in and 16 million out; the 95 percent on the cheap model costs about $16 and the 5 percent on the frontier model $64, so about $80. Review support is 4,000 summaries averaging 3,000 input and 1,500 output, 12 million in and 6 million out; the cheap half costs about $3 and the frontier half $210. The heartbeat is 30,000 small checks averaging 2,000 input and 100 output tokens on the cheap model, 60 million in and 3 million out, about $11. Retries add 4 percent, about $12. The monthly bill at the routed mix is roughly $320. The same workload with every call on the frontier model is about $2,900. The routing decision is worth about $2,600 a month, and it is one table in one config file that one person owns.

The same topology with a careless heartbeat — all four agents re-reading a 300,000-token context and producing 2,000 billed output tokens every five minutes on Fable 5.1, at $3.10 a check — runs 34,560 checks in a thirty-day month and about $107,000 in heartbeat spend alone, more than three hundred times the routed bill. That is why the heartbeat’s interval and model are budget decisions, not configuration details. The FinOps figure — 98 percent now managing AI spend — describes organizations discovering this arithmetic under deadline. The orchestrator who has already run it once has the only version of the number that survives contact with a budget committee: the itemized one, with the routing column explicit.