My token use grew 15x. That part wasn't me.

What the model produces per response · by week

And how often I prompt · per week

Chapter one: the tool got more capable.

The fair test is how much the model produces in a single response, because that number does not depend on how often I prompt. It sat near 120 tokens per response through February and March, then rose about fifteen times over as Opus 4.7 and Opus 4.8 arrived. Through that whole stretch my prompting rate stayed roughly flat, which is what made the original conclusion easy. The growth traced to the tool.

Chapter two: then I started using it more.

Output per response peaked in late May and has settled between roughly 1,000 and 1,600 since. My prompting rate did the opposite. It went from about 850 prompts in the last week of May to more than 3,000 a week in late June and July, roughly a fourfold jump. The first wave of growth came from the model. This one is me.

Fable 5 first appears in my logs on 9 June, goes quiet two days later, and does not return until 2 July. That three week gap is the model being pulled from availability, not a change in how I worked.

Replies take longer now. The model is doing more, not running slower.

Median time from prompt to finished reply roughly tripled, from about 12 seconds in early March to the low 40s now. The first chart shows why. Each response carries far more output than it used to. The model produces words faster than before, and more words produced faster still add up to a longer wait.

Median time per reply · by week

How I worked this out

Claude Code logs every session and records the token counts the service reports for each reply. I added up 23 weeks of those logs, from 16 February to 23 July 2026, across every folder I have ever run Claude Code in. The figures come from the tool's own records, with nothing inferred. Output per response is tokens produced divided by the number of times the model acted, which cancels out how often I prompt. Time per reply is measured from my prompt to the last thing the model wrote before I typed again.

One correction worth naming. A single reply is written to the log once per content block, so the same response can appear two or three times. The first version of this analysis counted those repeats. This version removes them by unique message id, which is why the numbers here run lower than the ones I published in May. The shape of the trend is unchanged.

The final bar is a partial week, 20 to 23 July.