I ran $24,000 of Claude through my terminal in August. Here is what it built.

In July, Anthropic invited me into the Claude for Open Source Program. The deal is simple: six months of Claude Max 20x, free, for maintainers of open-source projects they want to support. I have been building the Iris web framework for Go since 2016, so I said yes before I finished reading the email. The screenshot of that invitation is pinned on my GitHub profile, and I posted about it on X the same day.

Max 20x is the largest subscription Anthropic sells to individuals. It's a flat monthly fee with the highest rate limits, meant for people who live inside Claude Code all day. I wanted to know what "all day" adds up to, so on September 2nd I ran ccusage over my local Claude Code logs for August and exported the report.

There is one more thing you need to know to read the numbers right. August is not a working month in Greece. It is barely a month. Athens empties out, the bakeries close, and the only people left are tourists and the one Greek who forgot to book a ferry. This year that Greek was me. My coworkers were on the islands. My friends were sending photos of sunsets with captions like "you should be here". I was at my desk in Athens, in front of a terminal, replying "one sec, running tests". Every one of the 29 days below is a day I chose the keyboard over the sea, and I still can't decide whether that is a confession or a brag. My mother has decided. It's a confession.

One thing to be clear about before the numbers. ccusage prices every token at the public API list price. The figure below is what August would have cost if I had paid per token through the API. It is not a bill. The subscription is flat, and the program covers it. That gap is the whole point of this post.

The cache line is the one that surprised people when I showed them the report. Out of 31 billion tokens, only 1.5 million were fresh input that I typed or attached. Everything else was Claude re-reading the same repositories, the same CLAUDE.md files, the same skills, hour after hour, and prompt caching kept the price of that re-reading down to a fraction. Without caching, this month would have cost several times more, and it would have been much slower.

The report came from a single command:

npx ccusage@latest claude daily --since 2026-08-01 --until 2026-08-31 --json > august_metrics.json

Opus 5 did most of the work. It ran the long refactors, the test writing, the multi-file migrations. Fable 5 is the more expensive model, and I used it on purpose for the parts where a wrong design decision would cost me a week: API shape, concurrency review, security passes. Almost a third of the spend went there. That is a lot, and I would do it again. Sonnet 5 handled the subagent chores, scripted edits and searches. Haiku showed up for a handful of tiny lookups and cost $1.29 for the whole month, which in Athens no longer buys you a freddo espresso.

Iris has 25,500 stars on GitHub, and its last public release is still v12. Version 14 has been growing in a private repository for months, and August was the month it took its final shape: 351 commits in 29 days. The module path is already github.com/kataras/iris/v14, on Go 1.27.

Server-Sent Events landed as a first-class package on August 19th: 51 files, 8,268 lines including the tests. The compress middleware learned zstd the same day. The day after, I merged a refactor of JSON binding and response handling that touched 223 files, the kind of change I used to postpone for months because of how tedious it is to keep every example and every doc page in sync. There is a new ai package for talking to model backends from an Iris handler, because that is what people build in 2026. And v14 ships with something I have never shipped before: a Claude Code plugin and an Iris skill inside the repository itself. About 7,900 lines of structured documentation, so anyone running Claude Code on an Iris project gets a model that already knows the framework's conventions.

Earlier in the month came the less glamorous passes. A handler chain safety and security fix across 123 files. Numeric conversion and counter concurrency bugs. A golangci-lint job in CI. A docs-drift hook that complains when the book falls behind the code.

The new Iris website and the e-book moved in step. The book got chapters on AI backends and SSE, the framework comparison page grew, and every code sample switched to the v14 binding API and Go 1.27.

Two smaller libraries got real attention too. kataras/pg, my PostgreSQL layer, had a security and correctness pass, a thread-safe LISTEN/NOTIFY listener, and a retry loop that finally respects context cancellation before it sleeps. kataras/jwt shipped v0.2.0 with blocklist concurrency fixes, moved to encoding/json/v2, and fixed EC public key parsing. And server-benchmarks, my HTTP server benchmarking tool, was rewritten from scratch with a modular architecture and current framework versions.

Here is the honest accounting. Iris v14 got 351 commits in August, and it was not the only thing on my screen. A new product I have not announced got a lot of them too, and I am not putting a number on that one yet. It's an AI product. I will write about it when it's ready. Not before. Yes, I know that is exactly what everyone with an unannounced AI product says. Mine is different, which is also exactly what they say.

The number that matters to me is not the dollar figure. It's the 29 consecutive days, including August 15th, the biggest holiday of the Greek summer, when every church bell in the country was ringing, every taverna on every island was full, and I generated 5.76 million output tokens. My peak day was a Sunday. I am aware of how that sounds. Claude Code became the way I work, not a tool I reach for.

A few things made the difference. I plan before I let it edit: plan mode first, a written plan file, then implementation. I keep a CLAUDE.md with hard rules, and the hardest one is that Claude never commits. It stages, it writes the commit message, and I press the button. A hook enforces that, so I don't have to rely on trust. I write skills for repeated work, like the docs-drift check that keeps the Iris book aligned with the code. I use subagents for code review, with a second model reading what the first one wrote.

The limits are real. On August 18th Claude generated 7.95 million output tokens for me in a single day. No human reads that much code. I read the diffs that mattered, the tests covered a lot of the rest, and some of it I will find out about later. It still needs a maintainer who knows the codebase well enough to say no. Now and then it produces a clean, plausible design that is wrong for reasons only ten years of issue reports would teach you. What it gives me is the ability to carry a framework, a book, a website, three libraries and a product at the same time. I could not do that in June.

I've been maintaining open source for a decade, and this is the first time a company has handed a maintainer the exact tool they needed, at a scale that changes what gets shipped. Six months of Max 20x, no strings attached. Nobody asked me to write this post, and nobody at Anthropic knows what a freddo espresso costs. Thank you.

If you maintain an open-source project, look up the Claude for Open Source Program. If you use Iris, v14 is close. If you are one of my friends reading this from a beach: I saw the photos. All of them.

Reproduce the report with npx ccusage@latest claude daily --since 2026-08-01 --until 2026-08-31. My repositories live at github.com/kataras. The July announcement is at: @MakisMaropoulos/status/2028949824230371798.