I pay a flat monthly subscription for Claude. That price does not move when my usage does, so I have never actually seen what my habit costs. I priced six months of my own logs at published API rates to find out.
In February the metered figure would have been about twice the subscription. By June it reached twenty-five times, and July is tracking higher again. The subscription price never moved. My usage grew into it, then straight past it.
This is the part a flat rate hides. Nothing in my workflow tells me a given afternoon was expensive, so nothing in my workflow discourages the afternoon. That is the point of the pricing model, and it is worth seeing plainly at least once.
It is what these exact sessions would have cost billed per token at list prices. It is not a forecast of what I would spend on a metered plan. Priced by the token I would work differently, and almost certainly less, long before reaching six thousand dollars in a month. The number measures the value of the flat rate, not a bill I avoided.
Claude Code records the token counts the service reports for every reply. I summed those from 16 February to 23 July 2026 and priced each category at published rates: fresh input at list, cache writes at 1.25× or 2× depending on lifetime, cache reads at 0.1×, and output at the output rate. Model mix is taken from the logs rather than assumed, so each reply is priced at the rate of the model that produced it.
Two honest caveats. Replies are written to the log once per content block, so a single response can appear more than once. I removed those repeats by unique message id. And retention differs across the window: before mid-June only my main working folder was preserved, while afterwards every folder is present. In the period where I can see all of them, that main folder accounts for roughly 92 percent of activity, so the earlier months are undercounted by something in the range of eight percent.
February covers 16 to 28 only, and July covers 1 to 23. Both are drawn lighter.