---
title: 'In the News: October 23, 2026 (Evening)'
description: "One developer's logs show 96.66% of tokens in a busy hour were cache reads, and 42% of seven months of agent spend went to agents managing agents."
canonical_url: 'https://darkfactory.dev/news/2026-10-23-evening'
markdown_url: 'https://darkfactory.dev/news/2026-10-23-evening.md'
collection: news
date_published: '2026-10-23T19:25:00-04:00'
date_modified: '2026-10-23T19:25:00-04:00'
---

# In the News: October 23, 2026 (Evening)


A developer who counted seven months of his own Claude Code and Codex logs reports that most of the spend went to the setup and the supervision around the work, not to the work. It is one person's data, but he published the script that lets you run the same count on yours.

## 1. A seven-month token count finds most spend in re-reading and agent supervision

**[Where the tokens go - the end of maxxing?](https://morkeeth.substack.com/p/where-the-tokens-go-the-end-of-maxxing)** · Morkeeth (independent developer) · Morkeeth on Substack, October 21, 2026

The author counted every token his Claude Code and Codex sessions logged from March 8 to October 18 and prices it at $33,807 on published list prices. He sorted the use by what the work was for and reports that 42% went to agents managing other agents and 15% to apps people can actually use. In one hour with 1,701 API messages he counts 284,146,451 tokens, of which 96.66% were cache reads and 0.50% new output.

Two measurements bear directly on how a factory is configured. His normal setup sent 41,466 tokens on the first model call, against 16,603 for a bare session; he attributes the extra 24,863 to instructions, tools, skills, plugins, hooks and connected servers. And from September 18 to October 18, his scheduled-task runs made 2,679 model calls across 425 polling periods at $588 list price, or $1.38 per run. He notes that carrying a long conversation into a one-line status check is expensive. His fixes are to run checks as fresh one-shot sessions, keep one job per session, and trim what loads at startup.

He also reports that his own meter was wrong. For 1,763 matched Claude Code sessions, the harness recorded $17,991 and the token tool recorded $12,619; advisor calls explained $1,290 of the gap and $4,082 stays unexplained. He lists the limits himself: one setup, list-price estimates rather than invoices, and partly heuristic attribution of spend to projects. He also says Claude and Codex agents counted the logs and drafted much of the post. A reviewer he asked for comment says the dollar total should carry a range. As he puts it, "Token type is not a productivity score."

**Why it matters:** If you run scheduled or polling agents, the post shows a cost you can measure on your own logs: how large the first call is before any task text, and how much of the bill is cache reads. Treat the dollar totals as estimates, since the author's two counters disagreed, but the first-call size and the per-run polling cost are the numbers worth checking in your own factory.