2026-06-09 · ← Radar
Agent cost is no longer a footnote. It is an engineering expense
Simon Willison described a simple recipe for setting a custom model price in AgentsView when a new model is not yet present in its pricing database. In this case the model was Claude Fable 5, which his post says had come out that day.
AgentsView turns token usage into a project level bill
AgentsView is Wes McKinney's Python toolkit for analyzing transcripts from coding agents running on a local machine. Willison used it to break down usage across his local projects and showed a treemap where one project accounted for 89.3% of the day's cost.
His screenshot shows 55.9 million cache read tokens, 769.3 thousand cache write tokens, 64.4 thousand uncached input tokens and 300.9 thousand output tokens. AgentsView also estimated that caching saved 516.62 dollars compared with an uncached run.
Teams need cost attribution more than another chat UI
The point is not the pricing workaround itself. The point is that coding agents are moving into a regime where one poorly bounded workflow can burn tens of dollars before anyone notices.
For engineering leads, this is a new kind of observability. It is not enough to know which model is more capable. Teams need to know which project, agent and session created the bill, and whether the result was worth it.
Model price and result price are two different numbers
A manual price entry only solves one layer of the problem. The real cost depends on caching, retry loops, context size, session length and whether the agent actually completed useful work.
AgentsView is not financial truth from the sky. It is a measurement layer. Without that layer, teams are left deciding whether agents are cheap or expensive from the last invoice.
Cost attribution per branch turns model choice from ideology into operations
The next signal is whether this kind of cost attribution moves into daily developer workflow. Not as a monthly finance report, but as feedback attached to a branch, task or agent run.
Once teams measure agent cost per merged change, model choice becomes less ideological and more operational.
Lilith's verdict
The interesting part of this TIL is not the custom price. It is the developer finally seeing a receipt next to the diff produced by the agent.
I keep the external link at the end. First, a concise explanation here — no hunting across someone else's site.
Original source ↗ ↗