Codex cost per project: how to measure it Codex session files record each session's working directory, so per-project cost can be calculated by grouping sessions by that path and multiplying input, cached input, output and reasoning tokens by their respective API rates, according to a UsageAtlas guide. The guide warns that directory attribution is imperfect — a session that starts in one repository and edits another still belongs to the first — and that on a subscription the marginal cost is zero until a limit is hit, so API rates serve only as a shadow price. It recommends normalizing cost by merged pull requests or shipped features rather than hours, tracking a project for a full week before ranking it, and recording the model used, since a model switch can double a line without any behavior change. Quick answer Codex session files record the working directory, so cost per project is a two-step calculation: total tokens for that directory, then multiply by the rate for each token type. On a subscription the marginal cost is zero until you hit a limit, so value the tokens at API rates if you want a comparable number. Attribute sessions to a project 1. Open the session file and read the recorded working directory. 2. Group sessions by that path, not by the repository name; monorepos and worktrees share a name. 3. Total input, cached input, output, and reasoning separately. 4. Exclude sessions that only read a file from the project while working elsewhere. Directory attribution is imperfect. A session that starts in one repo and edits another still belongs to the first. Note that when a project looks unusually expensive. Convert tokens to money cost = input tokens input rate + cached tokens cached rate + output tokens output rate | Token type | Why it has its own rate | Common mistake | |---|---|---| | Input | Fresh context sent to the model | Using output rate for everything | | Cached input | Discounted reuse of a stable prefix | Counting it as full input | | Output | Generated text | Forgetting reasoning tokens | | Reasoning | Billed with output, often shown apart | Double counting | Read the rates from the provider's current pricing page. Do not hard-code a rate you saw in a comparison post. Compare projects fairly - Normalise by merged pull requests or shipped features, not by hours. - Separate exploration from implementation on the first pass. - Track the same project for a full week before ranking it. - Record the model used; a model switch can double a line without any behaviour change. UsageAtlas https://usageatlas.com/ download keeps the directory and model fields together, which is enough to produce this table without a spreadsheet. Attribute sessions before you divide Per-project cost only works if each session is attributed to one repository. Codex records the working directory per session, so the attribution is available; the mistake is dividing a monthly total by the number of projects you touched. - Group sessions by working directory. - Exclude sessions with no completed work. - Keep shared setup work in its own bucket. - Separate model cost from plan cost. UsageAtlas https://usageatlas.com/ download groups local sessions by project and shows the 30 or 90-day total, which is the number to divide. The formula is in AI coding cost per project https://usageatlas.com/guides/ai-coding-cost-per-project . Questions, answered Does Codex store the project name? It stores the working directory. The project name is derived from that path, and a monorepo will collapse several projects into one. Is cost per project meaningful on a subscription? Only as a shadow price. You do not pay per token until you exceed the allowance, but valuing usage at API rates shows which projects would be expensive if you did. Should I include cached tokens in project cost? Yes, at the cached rate. They are still billed, just cheaper, and ignoring them understates the total.