The Orchestrator

The two cases I want to examine both happened in the summer of 2025, both involved coding agents, and neither involved anyone doing anything wrong. That is why I chose them: a cost surprise that arrives with everyone behaving reasonably is a property of the system.

On June 16, 2025, Cursor changed how its $20-a-month Pro plan was priced. The old plan metered requests — five hundred a month, some models counting double. The new plan replaced the count with twenty dollars of frontier-model usage at API prices, plus unlimited use of an “Auto” mode that routes to whichever frontier model has capacity. Eighteen days later the company apologized: the changes “were not communicated clearly,” the included usage had been described as “rate limits” when it was a credit pool, and anyone with “a surprise usage bill” between June 16 and July 4 would be refunded in full. The stated reason for the change is the sentence that matters here: “New models can spend more tokens per request on longer-horizon tasks. Though most users’ costs have stayed fairly constant, the hardest requests cost an order of magnitude more than simple ones.”¹

Read that as an architectural statement before reading it as a pricing one. A request was a reasonable unit of account when a request was bounded: one question, one answer. Agentic work broke the unit. A single request could now run for many minutes, read a large slice of a repository, call tools, retry, and consume ten times the tokens of its neighbor, and a price denominated in requests was quietly subsidizing the long tail from the short head. Cursor moved the price onto the thing that varied, and the surprise bill was the variance in users’ own workloads becoming visible for the first time. The apology is the company’s own account; it does not say how many customers were affected or how large the bills were, so it establishes the mechanism and not the magnitude.

The second case is the always-on tax landing on a vendor’s bill instead of a customer’s. On July 28, 2025, Anthropic told subscribers to its Pro and Max plans that weekly usage limits would take effect on August 28, on top of the five-hour limits already in place, because some subscribers were running Claude Code “continuously in the background, 24/7” — alongside account-sharing and resale that violated its usage policy. The company estimated the limits would touch fewer than 5 percent of subscribers and told the heaviest Max users they could buy usage beyond the limit at standard API rates. TechCrunch, reporting the change, noted that Anthropic had described itself as constrained on compute and counted at least seven partial or major Claude Code outages on the company’s status page in the preceding month.²

Here the surprise ran the other direction. A flat monthly price is a bet that the median customer stays near the median. An always-on agent is a heartbeat that never stops, and enough of them turned the cost of serving the plan into the thing that no longer fit the price. Anthropic’s answer was a limit, and then a price — API rates — for anyone who wanted to continue past it: the same two levers this chapter hands to the orchestrator, applied by the one party who could see the whole workload. What the announcement does not establish is how much those subscribers consumed; under 5 percent is the only number in the record.

Set the two beside each other and they say one thing. When the unit of work changes shape — from a request to a long-horizon task, from a session to a heartbeat — the bill changes shape with it, and whoever is still pricing the old shape gets the surprise. In 2025 that was vendors and their individual customers. The systems this book describes move the same variance inside an organization, where the party pricing the old shape is finance and the party that changed it is you.