LLM cost management for AI teams.

LLM cost management is the practice of monitoring, attributing, controlling, and optimizing an organization's spending on large language models. The spend itself has two halves: the model calls your product makes on a customer's behalf, and the AI a person at your own company drives to get their work done.

It should, and most tools only cover one half. Gateways and observability see only traffic routed through them, so a Cursor or Codex session is invisible to both. Reading billing and usage data instead reaches the spend your product makes and the spend your own team drives alike, down to what each merged pull request cost. Those agents are not only an engineering line item either: anyone at your company who ships work through them shows up the same way.

Neither label covers all of it, which is why the useful split is who the AI was working for. Calls your product makes to serve a customer behave like cost of goods sold and set your AI gross margin. Spend your own team drives is operating expense, but not always R&D: an engineer shipping a feature is, a designer or a support lead doing day-to-day work is not. Dev and staging traffic is neither, which is one reason environment is a dimension you group by.

Cloud resources carry tags you set at provisioning time, so allocation is mostly a tagging discipline. A model call carries only what the provider records: a project, an API key, a model. None of that knows your customer, your feature, or your repository, and most teams pay four or five vendors at once.

It depends where the figure came from, so a serious tool says which. Spend read from a billing API is the billed number; spend computed from token counts is an estimate of it. SuperPenguin shows the two side by side per provider and names the gap rather than smoothing it.

Both, which is why it usually falls through the cracks. Engineering controls what moves the number, meaning the model, the prompt, and the retry policy. Finance owns the budget and the invoice. It works when one figure an engineer can trace to a request is also one a finance reader can trace to the bill.