Coding agents are where AI labs found product-market fit

Simon Willison's latest essay makes a claim that has been quietly true for a while: Anthropic and OpenAI have found product-market fit, but in coding agents and enterprise deployments, not consumer chat. The numbers he uses to ground it are specific. He himself burned about $2,180 of API token value last month while paying $200 in subscriptions, which is the kind of usage gap that only closes when somebody starts charging by the token. Both labs did exactly that in April, switching enterprise plans to API-based pricing, and Willison reads the new Claude tier as roughly 1.4x more expensive than the predecessor once you adjust for tokenization changes.

The macro picture lines up. He points to the SpaceX SEC filing that names a $1.25 billion-per-month agreement with Anthropic through May 2029 as evidence of the inference bill these workloads generate. Anthropic's projected $10.9 billion Q2, up from the $4 billion run-rate that was concentrated in Cursor and GitHub Copilot, fits the same story. Willison adds a less obvious tell: OpenAI is hiring for 703 roles with about 33% tagged enterprise, Anthropic for 390 with about 27% enterprise. Sales orgs do not grow that way to serve free-tier users.

His framing is that coding agents hit the rare combination of being expensive to run, valuable enough that high-salary developers do not flinch at API pricing, and token-hungry in a way that chat is not. That is what turns inference costs from a problem into a revenue model.

Why it matters

If you build on Anthropic or OpenAI APIs, this is the pricing trajectory to plan for: usage-aligned, growing more aggressive at the enterprise tier, and pointed at coding workloads first. For internal tooling teams, that means the per-developer math on agent usage is about to look more like a real line item than a hobby budget, and procurement will start asking.