Last week, an engineer at a well-funded agent-platform startup shared a screenshot of his company's usage dashboard in a conversation with us. It was a good dashboard — clean, honest, itemized per agent. It showed a small fleet of five or six agents spending $803 in a single week. And it showed something else, in a quiet corner of the cost breakdown: of every dollar spent, about 99 cents went to LLM tokens. Compute was under a penny. Storage was zero.

The dashboard was offered as evidence of cost control. Look — we have budgeting built in. We can see everything.

But a dashboard measures spend. It does not shrink it. Knowing that one agent burned $661 on 7.3 million tokens is not the same as spending less. A speedometer is not a brake.

The elephant

Here is what nearly everyone in the room is not looking at: a large share of those tokens are not doing intelligence work. They are doing clerical work at intelligence prices.

When an agent reads the web the default way, it feeds raw HTML into the model — navigation bars, cookie banners, script tags, tracking pixels, footers, all of it — and pays token rates for every character. Our measured median across real pages: raw HTML is about four times larger than the clean article inside it. That means roughly three-quarters of every web-reading token an agent buys is spent on content nobody, human or machine, wanted.

Fetching a page is deterministic. Stripping the boilerplate is deterministic. Parsing a table is deterministic. None of it needs a model. All of it is being billed as if it does. That is the elephant — it is massive, it is standing in the middle of the usage dashboard, and the industry's attention is pointed at the other side of the room, debating which LLM to switch to.

What the numbers look like

We ran an executive-level estimate on that startup's dashboard (anonymized here; the numbers are theirs, the assumptions are ours and stated plainly).

Clean reading alone: for assistant- and research-style agents, a conservative 25-35% of input tokens are raw web content. Cleaning removes about three-quarters of those tokens. That is roughly a 20% cut to the total LLM bill — about $690 a month on their current fleet — and the reads that produce it cost $0.002 each. The savings-to-cost ratio is around 100 to one.

A full deterministic layer: clean reading plus four non-competing pieces of boring infrastructure — prompt caching, a search API, a retrieval store, and a code sandbox for math and data work. Each attacks a different slice of token waste. Combined, an honest range is 45-60% of the LLM bill. On their fleet: roughly ,700 a month. And that dashboard showed five agents. The company talks publicly about thousand-agent teams. At that scale this is not a line item; it is a seed round.

A 55-minute proof

The counter-example, from our own world: a founder we work with runs an autonomous agent that does real prospecting — reading company pages, regulatory filings, and news, hundreds of pages per session. A common rule of thumb for an agent doing heavy web work the default way is on the order of a dollar a minute in tokens. His agent recently worked for 55 minutes straight.

The session cost $5.80.

Not because the model was cheaper, but because the model almost never saw the raw web. The agent paid a fifth of a cent per page to read clean markdown, and the expensive part of the stack — the intelligence — was reserved for the part of the job that actually required intelligence. That is the whole thesis in one receipt.

The Skim Audit: free, no signup, no wallet

The estimate we ran on that startup's dashboard took us an afternoon and two screenshots. It is worth real money to the company in the screenshots. So we are making it a standing offer.

Send us your agent usage dashboard — a screenshot is enough — and we will send back a free, executive-level savings estimate: where your token spend is doing deterministic work at model rates, what routing it to flat-priced infrastructure would save, and which changes pay back first. Stated assumptions, honest ranges, no black box. If the answer is "your spend is already lean," we will say that too.

No signup. No wallet. No sales call unless you ask for one. Your numbers stay private — we never publish or reference them without your permission.

The details are on the Skim Audit page. The short version: email hello@skim402.com with the subject "Skim Audit" and attach your dashboard.

The elephant does not care whether you look at it. It eats the same either way. But it is remarkable how much smaller it gets once someone finally points at it.