A record, not a boast · 15 January to 6 July 2026
One person, no engineering background, running Claude Code most days for six months. This is what the meter actually read, what the tokens actually bought, and what broke along the way. Every number below is measured, not remembered.
On 15 January 2026 I started logging every Claude Code session. By 6 July the counter stood at 59,234,351,520 tokens across 784 sessions: 173 days on record, including a 31-day unbroken streak that was still running when this page shipped. The counter is public and updates daily at the dashboard, with the raw numbers at data.json.
As of publication: 59.2B tokens and counting.If you are going to make claims about working with AI, put a meter on it first.
Fifty-nine billion sounds like generation on an industrial scale. Mostly it is not. The bulk of the volume is cache reads: the model re-reading the working context of long sessions, the same files and conversation loaded again and again as work continues. That is the token equivalent of an employee's working memory, not their typed output. Fresh generation is a small slice of the total.
The record is honest about its own gaps too. The first weeks predate proper instrumentation, so 19.1% of the history (33 early days) is spread from weekly and monthly archive totals rather than measured per-day. The dashboard marks that estimate zone visibly rather than smoothing it away, and the measured share only grows: history is never rewritten. A second, independently coded tracker over the same raw sources agrees with the headline to within 2.1%.
A number you cannot audit is an anecdote with confidence.
Tokens are the fuel, not the point. The point is the launch log, which is also public. Over the same 173 days: 90 launch-log entries across the portfolio, 4,627 human commits, and 114 million generated words, roughly 1,427 novels of output, most of it code, configuration and drafts rather than prose anyone reads.
Two examples with real stakes. Beat the Gaffer World Cup is a predictions game with a live user base playing through the 2026 World Cup right now; it has been audited, patched and improved throughout the live tournament. And the usage dashboard itself is a self-documenting build: the thing measuring the tokens was built with the tokens.
The launch log is the CV. The token counter is just its receipts.
The six months were not a flat line. The biggest single day was 5 May 2026 at 2.85 billion tokens; 7 June, the World Cup game's launch day, ran it close at 2.78 billion. Weeks classify cleanly into sprints and maintenance: a launch week runs multiples of the trailing median, then the line drops back while real users, cron jobs and monitoring take over.
That shape matters more than the total. It is the difference between a tool you demo and a tool you operate. The quiet weeks are where the automated pipelines, scheduled agents and watchdogs earn their keep, because output continues while attention is elsewhere.
Sprints ship products. Maintenance keeps them true. The meter sees both.
Actual spend over the six months: $1,200 in subscription fees. The same token volume priced at API rates would be $141,797, a 118x multiple. That figure is an equivalence, not a saving; nobody would have bought those tokens at list price, and the dashboard labels it exactly that way. The more grounded framing: at my day rate of £1,500, the six months of shipped output equates to a contractor engagement that would price in six figures, delivered alongside everything else, for the cost of a phone contract.
Averages make it concrete: 12.3 million tokens per human commit, 658 million per launch. Those are absurd numbers by hand-tool standards, and that is the point. Token volume is what it costs to give a machine enough context to do real work on a real codebase. The economics only look strange until you stop pricing the fuel and start pricing the freight.
Stop pricing the fuel. Price the freight.
This section is why you can trust the other seven. A six-month record with no failures in it is a brochure.
Instruments drift. The discipline is building the gauge that catches your own gauge lying.
The durable lesson of the six months is not a prompt technique. It is that working with AI at this volume turns one person into a very small institution, and institutions need infrastructure. Everything that broke above produced a rule, and the rules compound: verified numbers only, never from memory; every deploy gated on contrast and mobile checks; regression gates on anything a pipeline publishes; one writing session per repository at a time; and a persistent memory system so the five-hundredth session starts smarter than the first.
None of that is glamorous. All of it is why the pace in section three is sustainable rather than a burst. The instructions became the infrastructure.
The prompts are disposable. The rules are the asset.
One person plus this tooling now ships at a rate that used to need a small team. That is not a claim about everyone; it is a record of one case, with the receipts published and the failures itemised. The counter is on pace to cross 100 billion around 29 August 2026, and when it does, the write-up will use the same rule this one did: every number from the data file, none from memory.
Asked for a one-line verdict on the six months, here is mine, unedited: I've learnt a whole new way of working and I think it's here to stay; there's no going back now.
If your organisation wants to know what this looks like from the inside, the dashboard is public, the method transfers, and I am reachable at hello@mikelitman.me.
Measured, not remembered. That is the whole method, and it fits in three words.