Iris maintainer logs $24,000 of API-equivalent Claude usage in one month

A month of agent usage, priced

The developer behind Iris, a Go web framework with roughly 25,500 GitHub stars, has published a detailed account of what a free Claude Max grant from Anthropic produces in the hands of a heavy user. Writing on dev.to, the maintainer — posting under the handle kataras — says Anthropic invited him into its Claude for Open Source Program in July, offering six months of Claude Max 20x at no cost. Per the post, Max 20x is the largest subscription Anthropic sells to individuals: a flat monthly fee with the highest rate limits, aimed at people who work inside Claude Code all day.

To measure what that adds up to, he ran the ccusage tool over his local Claude Code logs for August. The headline figure is $23,973.90 — but, as he is careful to note, that is not a bill. ccusage prices every token at public API list prices, so the number represents what the month would have cost if paid per token. The subscription is flat, and the program covers it. That gap between metered and flat-rate pricing is, he writes, the point of the exercise.

The numbers behind $24,000

According to the exported report, August included 29 consecutive active days, averaging $826.69 of API-equivalent spend per day. Total throughput was 31.17 billion tokens, of which 30.61 billion — 98.2% — were served from cache. Fresh input, the tokens he actually typed or attached, came to just 1.54 million. Output totalled 91.36 million tokens, about 3.15 million a day. His peak day was a Sunday, August 30, at $1,351.85 and 2.02 billion tokens; his biggest output day, August 18, produced 7.95 million output tokens.

The cache figure is the standout. Nearly all token traffic was the model re-reading the same repositories, CLAUDE.md files and skills, hour after hour, with prompt caching keeping the effective cost to a fraction of list price. Without caching, he writes, the month would have cost several times more and run much slower.

Model usage split as follows: Opus 5 carried 68.7% of spend ($16,474.82) on long refactors, test writing and multi-file migrations. Fable 5, described as the more expensive model, took 29.0% ($6,949.54) and was reserved for decisions where a wrong call would cost a week — API shape, concurrency review, security passes. Sonnet 5 handled subagent chores at $526.85, and Haiku 4.5 logged $1.29 for the entire month.

What it shipped

The output went chiefly into Iris v14, which has been growing in a private repository while the last public release remains v12. August added 351 commits in 29 days, targeting the module path github.com/kataras/iris/v14 on Go 1.27. Highlights per the post include a first-class Server-Sent Events package (51 files, 8,268 lines including tests), zstd support in the compress middleware, a JSON binding and response refactor touching 223 files, and a new ai package for calling model backends from an Iris handler. V14 also ships a Claude Code plugin and an Iris skill — roughly 7,900 lines of structured documentation — so agents working on Iris projects start with the framework's conventions already loaded.

Less visible work included a handler-chain safety and security fix across 123 files, numeric conversion and counter concurrency bug fixes, a lint job in CI, and a docs-drift hook that flags when documentation falls behind the code. The side libraries kataras/pg, kataras/jwt (v0.2.0) and server-benchmarks all received attention, and the author says an unannounced AI product also consumed commits, without giving a number.

Guardrails and limits

His workflow leans on structure: plan mode before edits, a written plan file, and a CLAUDE.md with hard rules. The hardest rule is that Claude never commits — it stages changes and drafts the commit message, while a hook enforces that a human presses the button. Skills handle repeated work, and a second model reviews what the first one wrote.

He is also candid about the ceiling. Nobody reads 7.95 million output tokens in a day; he reviews the diffs that matter, relies on tests for much of the rest, and expects to discover some of it later. The agent, he argues, still needs a maintainer experienced enough to say no to clean, plausible designs that are wrong for reasons only a decade of issue reports teaches.

Why it matters

This is a rare, quantified first-hand account of sustained agentic coding at production scale. It shows the economics that make the pattern feasible — cache-dominated throughput with human input as a rounding error — and what Anthropic's open-source grant program is designed to unlock: a solo maintainer simultaneously carrying a framework, a book, a website and several libraries, which he says was not possible before the grant. It also surfaces the open question. When generated output far exceeds what any human can read, review becomes triage, and the maintainer's judgement, not the model, becomes the scarce resource.