ttokunaga: $17,239.93 in AI use cost 36,300円

This post is a personal accounting exercise, but its structure makes it useful as a reference. The author gathered Codex usage from detailed logs plus an estimated database portion, and totaled Claude Code from 17 saved log sessions covering 5,404 API responses. Both sides were then priced with each provider's standard API rates as of October 3, 2026, using a comparison rate of 1 USD = 150 yen — a deliberately uniform yardstick rather than actual invoices.

What shapes the month is not only how much was used but where the money went. The author ran Codex on two accounts, 30,000 yen for personal use and 3,000 yen for work, and paid $22 for Claude Pro. He also notes that Claude Pro was 3,000 yen on Japan's App Store, about 600 yen less per month than what he had been paying on the web, and that most of his Claude volume was cache reads — about 2.073 billion tokens, or roughly 98.8% of the total. That mix is why a cache-read price cut lands so heavily in his figures.

The reading of the gap depends on the plans holding still. The article points to OpenAI's Pro 200 keeping its $200 price while its usage limits shrink after October 29, and to the new Pro 500 tier, which is why he is leaning toward moving some development to Claude Code. Whether the subscription-versus-API gap persists therefore hinges on these terms, and developers who bill the same way could be affected by exactly such changes.