Authors
Find out what your AI bill could drop to
No new hires, no new tools, no new policy. The AI line still went up, again. This is now the normal case, and the reasons are structural, not careless.
The pattern
CloudZero's AI Economics Pulse for October 2026, drawn from 422 of its customers' cloud bills, found AI's share of the median company's cloud spend has risen for 14 months in a row and is 4.5 times what it was a year earlier. In the same report, finance leaders guessed AI was 21 to 35 percent of their infrastructure spend. The bills said 3 percent. Both facts matter: the number keeps climbing, and the people responsible for it cannot see it clearly.
Five quiet reasons
1. Sessions get longer
A coding agent resends its whole history on every turn. As people trust the tools more, they keep sessions open longer and ask bigger questions. Each turn carries more context than the one before, so the cost per question rises even when the number of questions does not.
2. Agents multiply
One developer, one chat window was last year. Now it is one developer, three agent sessions and a review bot on every pull request. Automation runs when nobody is watching, which is exactly when nobody is watching the cost.
3. Defaults drift to the top model
Tools ship with the strongest model as the default, and nobody changes it back. Anthropic's price list puts Claude Opus 5.5 at 40 times Claude Haiku 5.5 per token on prompts up to 100k. Most of the work does not need the top model, but the default decides, not the task.
4. The cache goes cold
Anthropic bills a cache read at 5 percent of the input price, so repeated context is nearly free, while the cache is warm. Its default lifetime is five minutes, and any change earlier in the prompt rebuilds it at full price. Long pauses, reordered instructions and context compaction quietly push a team from paying 5 percent to paying 100 percent for the same text.
5. Nobody owns the number
With a shared account, the invoice is one line. Engineering says it is finance's problem, finance cannot say which team caused it, and the only lever anyone has is "use it less", which nobody wants to pull.
The fix is visibility first
Every reason above has a cheap, mechanical answer: trim the context each turn carries, route routine work to a cheaper model, keep the cache warm, batch the nightly jobs at half price. But none of them can be applied, or even chosen, until spend is visible by person, team, tool and model. That means a key per person and per agent, with a budget on each, and a log that records who used which model, how many tokens and what it cost.
Once that exists, the bill stops drifting because the drift has a name and an owner. This is the order we work in: a free review of one month's usage first, so the causes are measured rather than guessed, then a proxy tuned to the ones that actually matter for your team.
