Willison reconstructed the real economics of coding agents. In 2026 Anthropic and OpenAI moved their enterprise plans from flat fee to full API pricing, while frontier models cost more: GPT-5.5, Opus 4.7. Result: enterprise customers pay hundreds or thousands of dollars a month, and that number now moves with usage, not with a fixed subscription.
For anyone running an AI workflow in production, this breaks the quote. Flat fee let you budget a single line. Full API pricing turns the cost into a variable, tied to the tokens you burn every day. Add the infrastructure, the person who keeps the pipeline running, the refactor when the provider switches models, and the real annual TCO lands well above the API bill. The labs, for their part, sign billion-dollar compute contracts: Anthropic, 1.25 billion a month through 2029 according to SpaceX's S-1. That cost, sooner or later, flows downstream to the customer.
Why this matters for anyone building enterprise AI: the flat fee was a subsidy, the real price arrives now and has to be budgeted as a variable.