According to recent reports, I build a developer tool alone, with Claude Code and Codex running most of the day. For the last 30 days (8 October to 6 November 2026) I priced every call they made at API list rates.

The total: $8,360. Claude: $8,056. Codex: $304.

That isn't my bill. I'm on a flat blueprint, so what I actually paid is the plan fee. The $8,360 is what the same work would have cost through the API. I watch it because it tells me whether the plan is paying for itself, and which work is expensive.

In a recent development, A sevenfold spread between two days of the same person on the same projects. A monthly average hides that, and the busiest day is what a proposal's limits have to hold.

Recent findings suggest that about nine dollars in every ten went to Claude Opus. Everything else together β€” Fable, Sonnet, Haiku and OpenAI's GPT models through Codex β€” is the remaining tenth. If I want the number lower, choosing a model per task does more than anything else.

According to recent reports, in one October week, 45 pull requests merged on my main project. The agents' work on them came to $343.78, or $7.64 per pull request at list price. The most expensive single one cost $71.13. Seen that way, the useful question isn't the monthly total but which changes cost nine times the average, and why.

Claude Code and Codex both keep their own session logs on disk, and neither adds its sessions up with the other's. I read both into one local table and price every turn by model at list price. Nothing leaves the machine.

According to recent reports, the tool I use for this is the one I'm building: Caprock, a free, open-source (Apache-2.0) desktop app for macOS, Windows and Linux. You don't need it to do this, though. The logs are in ~/.claude/projects/ and ~/.codex/sessions/, and the prices are on Anthropic's and OpenAI's pricing pages. A script can do the same sum.

The full numbers, with the method, are on caprock.dev/numbers.

β†’ Read original article