What they did
The developer behind Apps by Mo builds about twenty Mac apps alone with Claude Code, and kept running out of the plan by midweek. After reading Claude Code's own usage logs on the Mac, they found reviews were the most expensive thing they did: before shipping anything that touches saving, sync or payments, Claude read the whole diff and hunted for problems. They split the job. Codex, on their ChatGPT plan, now does the bug hunt in read-only mode, and Claude Code only reads each finding and decides whether it is a real bug or a false alarm. Alongside that, long sessions end with a short handover note that a fresh session starts from, and each app keeps a small memory file with build commands and known traps that every session reads first.
How it went
Across 68 reviews, Claude used about 71% fewer tokens per review, and 120 of the 128 findings Codex raised were judged real bugs. Over one measured week, the same work came to about 280M tokens against an estimated 478M with Claude Code alone, which the author puts at 41% of the plan saved. The plan now lasts the week. They also found a reply late in a long session cost about 3.3 times one in a short session.
Worth knowing
The numbers come from the author's own reading of the logs, and the 478M baseline is an estimate of what the same work would have cost. The post ends by promoting the author's own Mac app, Atelier, which automates these steps; everything described can also be done by hand. You need both a Claude plan and a ChatGPT plan for the split.