*Every dollar figure here is an API-equivalent estimate at public list prices — a way to value the work, not a bill I actually paid. More on that in the methodology section.

We have one rule that never bends: you can’t improve what you don’t measure. We say it about ad spend, about content, about hiring. So it would be pretty hypocritical to pour thousands of dollars of AI into the business every month and have no idea where it goes.

So I instrumented it. Every Claude desktop session I run gets logged locally. Once a week a script reads every transcript, tallies the tokens by model, estimates the cost, and emails me a report. This article is that data, opened up for everyone.

If you’re an operator, a marketer, or a founder trying to figure out whether AI is a toy or a line item that deserves a budget, this is the honest picture of what heavy, daily use actually looks like.

New here? Start with our complete guide to how to use Claude, see how we use it to build marketing systems, meet the full agent roster, or watch a Claude agent build a website end to end.

📅 The last seven days

July 11–17, 2026 · 163 sessions · 7,122 assistant messages · 8.4M output tokens · ~$1,110.15

Here’s the week, day by day. Notice the shape — a quiet weekend, a steady climb, and then Thursday detonates.

The weekday-vs-weekend gap is the other tell. Friday the 11th cost $3.48. Thursday cost $431.60 — a 124× difference inside one week. Averaged out, each session ran about $6.81, but averages hide everything interesting here.

🧮 The all-time picture

2,009 sessions · 30.4M input + 95.5M output tokens · 18.35B cache-read tokens · ~$15,345.18

Zoom out and the numbers get large fast. Across every session on record I’ve generated 95.5 million output tokens — roughly the equivalent of writing 70+ million words. Input was another 30.4 million tokens. And then there’s the number almost nobody talks about: 18.35 billion cache-read tokens.

💸 Where the money goes: cost by model

One model eats more than half the budget.

This is the chart that changed how I think about model selection. A single model — Claude Opus 4.8 — accounts for 58% of all-time spend. The premium models (the Opus family) together are ~82% of the bill, while the fast, cheap models (Sonnet and Haiku) do a meaningful chunk of the work for pennies.

🔥 The busiest days on record

June ran hotter than July. The current week doesn’t even crack the top 10.

Context matters. This week’s $1,110 feels big until you line it up against the real peaks. My single most expensive day ever — June 14 — cost more by itself (~$1,099) than this entire week did.

Interesting wrinkle: day #8 (June 26) hit $549 with only 13 sessions, while day #7 (June 10) needed 100 sessions to reach a similar number. Session count and cost are only loosely correlated — a handful of deep, long-running agent jobs can outspend a hundred quick ones.

🛠️ What the AI is actually doing

Tokens tell you the volume. Tool calls tell you the work.

The most revealing dataset isn’t tokens or dollars — it’s which tools the agent reaches for. This is the fingerprint of real operational work, not chatting. Here are the top eight, all-time:

Read that top line again: the single most-used tool is browser automation — running JavaScript inside Chrome, 7,104 times, plus 4,310 screen-control actions and 1,800 navigations. Add it up and browser control is by far the dominant activity. The AI isn’t just answering questions; it’s driving software — logging into dashboards, pulling reports, filling forms, checking rankings. These are the same AI agents we deploy for clients, each with a defined role in the agent roster.

Behind that: shell commands (5,500+ combined) for running code and data jobs, file reads and writes (3,700+) for producing deliverables, and 1,660 web searches for live research. This is the profile of a digital worker, not a chatbot.

⚙️ How the tracking actually works

The whole system is four moving parts and runs itself.

You don’t need enterprise tooling to measure your own AI use. My entire pipeline is a local script and a scheduled job. Here’s the flow:

A scheduled job runs the analyzer every Monday at 8am. It regenerates the report, refreshes the dashboard, and an agent task emails me the summary an hour later. The only human step is reading it — and now, occasionally, publishing it.

🎯 Five takeaways you can steal

The bottom line

AI stopped being a novelty in our business the moment it became a measurable line item. $15,345 of API-equivalent work across 2,009 sessions isn’t a story about spending — it’s a story about leverage you can audit.

The tools you use every day are throwing off data. Capture it, chart it, and you’ll make better decisions in a week than you would guessing for a year. That’s true for your ad accounts, your funnels, and now your AI.