Cost, risk, and architecture for agentic AI in production. A newsletter for the people who answer for the budget.
Read the latest: You can't underwrite a bill you can't reproduce
Latest article
Run the same agent on the same task twice and you get two different bills. That variance, not capability, is what stalls agents in production: a program whose cost you can’t reproduce is one finance can’t underwrite. Gartner expects 40% of agentic projects canceled by 2027 over exactly this. The standard fix, a spend cap, doesn’t narrow the spread. It chops the tail off and leaves the run truncated, trading a cost you couldn’t predict for a failure you can’t either.
More from Stax
All articles →8 hrs ago
Half of US business now pays for AI, and adoption stalled the same quarter companies began capping their own staff’s access. Rationing is what it looks like when the bill outran the plan. The spend reduces to tokens, and there are exactly two levers on it: how much context rides on every call, and which model you put the work through. A staff quota touches neither.
1 day ago
The claim traces to one anonymous sentence in an Axios newsletter, and it tore across Reddit, X, and the news sites in days. It spread because every AI budget is under question right now and a giant dumb number is exactly what a skeptic wanted. To bill $500 million in a month you would have to generate 20 trillion tokens, about 7.7 million every second, nonstop, all month. The real overruns are smaller, named, and already on someone’s invoice.
3 days ago
Opus 4.8 put an effort slider in front of every Claude user, and it defaults to high. Anthropic’s own help text says the lower setting uses up rate limits more slowly. Read that backwards. The best research lab on agent cost keeps setting every default to the expensive end of its own curve.