The AI coding spend you can't see

Open your provider's billing page and try to answer one question: what did Tuesday afternoon cost you? If the honest answer is "no idea," you're not careless — the spend is genuinely invisible. Here's where it hides.

You never chose to send most of what you're billed for

The mental model most of us carry is "I pay for my prompts and the model's answers." The invoice disagrees. In a typical coding-agent session, the prompt you typed is a rounding error. The bulk of every request is context: the conversation so far, files the agent read, and the output of every tool it ran — shipped to the model again on every single turn.

• Re-read files. The agent opens the same unchanged file five times across a session, and you pay for it five times.
• Tool output nobody reads. A test run prints 4,000 lines; the three that matter are the failures. All 4,000 go to the model.
• History that never leaves. Turn 40 still carries turn 3's exploration, long after it stopped being relevant.

None of this appears on your screen. All of it appears on your bill.

Why the billing page can't help

Provider dashboards total tokens per day or per key. They can't tell you which session burned the money, which tool call was the whale, or — the question that actually matters — how much of it bought you nothing. Aggregates hide exactly the thing you need to see. We've written before about the team-scale version of this problem; it starts at one seat.

Step one: put a meter where the money flows

The fix starts with position. Your agent talks to the model through one pipe — put a meter on that pipe, on your own machine, and every request becomes visible: what was sent, what it cost, what was wasted. That's what TokenBeaver is — a local gateway between your agent (Claude Code, Cursor, anything OpenAI-compatible) and the model, with a dashboard served from 127.0.0.1. Your code and prompts never leave your machine.

Step two: stop the waste before it's billed

Seeing the leak is half the job. The gateway also slims each request before it's sent — nothing the model needed is lost, so you get the same answers from fewer tokens. In our internal testing across models that's come out to 40–70% lower API spend on the same day's work. Your number will be different, which is rather the point: the dashboard shows yours.

For tool-specific starting points, see cutting Claude Code costs and lowering your Cursor bill.

Estimated is good. Verified is better.

Free shows estimated savings — modeled locally from your own traffic, close enough to make decisions with. Pro verifies the same numbers against what actually happened in your sessions and adds the Activity Log: a searchable per-event history with filters, time ranges, and CSV export — the receipt you can put in front of anyone who asks where the AI budget went.