Chapter 3. Swarm Economics: When 15× the Tokens Pay Off
A spends roughly 15× the of an ordinary chat — and that is not a side effect but the very mechanism of its edge: per Anthropic, token spend alone explains 80% of the difference in quality. So swarm economics comes down to two questions: is the value of the task high enough to pay for the multiplier, and can you use engineering means — model tiering, , batching, and budgets — to cut its real cost several times over.
Where the 4×, 15×, and 90.2% Come From
The canonical source of the numbers is Anthropic's engineering post "How we built our multi-agent research system" (June 2025). Verbatim: " typically use about 4× more than chat interactions," and " use about 15× more tokens than chats." The price for it: an system with Claude Opus 4 as the lead and on Claude Sonnet 4 beat single-agent Opus 4 by 90.2% on Anthropic's internal research eval. On the BrowseComp , three factors explained 95% of the variance in quality — with tokens alone accounting for 80%, the rest being the number of and the choice of model. The authors' conclusion is honest: "multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance."
Three caveats, without which these numbers must not be cited:
- 15× is a measurement of a single internal system from mid-2025, not a law of nature. It was taken before 1M-token context models, cheap cache-reads, and server-side ; no vendor had published an updated multiplier by mid-2026.
- The 90.2% comparison was not token-budget-matched. Tran & Kiela (arXiv 2604.02460, April 2026) show that at equal budgets of "thinking" tokens, a single agent on reasoning consistently catches up with or overtakes multi-agent systems. A large share of a swarm's advantage is simply buying extra tokens unavailable to a single .
- Parallelism buys , not money. Anthropic reports cutting research time "by up to 90%" — while token spend grows ~15×. A practical example: parallel Claude Code subagents exhausted the Pro plan's window in ~15 minutes versus ~30 minutes sequentially — the same work, the quota burned twice as fast (dev.to).
The 2026 Price List: Raw Material for Tiering
economics is set not by the average price but by the price spread within a lineup — that is exactly what lets you pay for intelligence only where it is needed. prices as of July 2026, $ per 1M input/output tokens (Anthropic, OpenAI, Google):
This is where the open part of the chapter ends. The rest opens with a free registration — no payment, just an email address.