Claude Code cost per month: $5,898 on a $200 plan

Claude Code costs $20 a month on Pro, $100 or $200 on Max, or whatever your tokens come to on the API. What my own use was worth at API list prices is a different number: $6,056 in July, $5,898 in August, $1,764 in September.

None of those is a bill. They are what the same tokens would have cost billed through the API — and I paid $200 a month for Max, so August alone was 29 times the fee. Every plan's price, and how to choose between them, is in Claude Code pricing.

My months at list price

July starts on the 18th, when the record begins, and I used Claude Code on all 14 of its days. August is the one full month: 54 sessions on 29 days, $203 a day on the days I worked. All 75,884 replies included a model and token counts, so I could price every one.

That is one heavy user's machine, about fifteen times the $13 a day Anthropic gives as its average across enterprise deployments, from its cost page. It is not a forecast for anyone else.

Claude Code itself can no longer say what July cost: every July transcript has been deleted from disk after 30 days. The July figure exists because Caprock read those files while they were still there.

September is small because I was using Codex

From 10 to 28 September there is not one Claude Code reply in the record. The work in those weeks went to Codex: 14,998 Codex replies on 17 of those 19 days, in the same database.

So $1,764 is eleven days of Claude Code at $160 a day, not a cheaper month. Corrected on 3 October: this first read $2,093. One session I forked on 2 September carried 827 of its parent's replies, and Caprock counted them in both; since 0.68.0 it prices a reply once, whichever transcript it appears in. A machine running two agents needs both counted; Codex has its own guide.

What a token costs: most of it is re-reading

79%

of July's list price was cache reads

78%

of August's

83%

of September's

Fresh input, the part of a prompt that is neither written to the cache nor read from it, came to $0.37 in July, $0.29 in August and $0.08 in September. Nearly everything Claude Code sends is written to the prompt cache once — 11% to 17% of each month — and then read back on every turn after it. Output, every line of code and prose Claude wrote, was 4.7% to 5.1%.

A cache read on Opus 5 costs $0.50 per million tokens, a tenth of fresh input. It is still the biggest line, because each turn re-reads the whole conversation: 10.15 billion cache-read tokens in July alone. What fills that conversation is the subject of where the money goes.

Plan or API

I paid $200 a month for Max through these months; I pay $100 now. Against that $200:

29×

August, the one full month: $5,898 of tokens for $200

30×

July, and that is only 14 days of it

9×

September, 11 days of Claude Code; the rest went to Codex

For me, the flat plan made that much work possible. It is not a saving I would have made on the API: without the flat fee I would have run less.

I did not run into the plan's limit either. From 20 August to the end of September, Claude Code ended 29 turns on an error and Caprock kept every one: 12 were a conversation longer than the model accepts, the rest access refusals, server errors and the like. Not one was a usage limit.

Run it on your machine

/usage shows the current session. For months, on 1 October:

$ caprock export events --agent claude --format csv --out claude.csv
258359 events rows → claude.csv
$ awk -F, '$6=="turn.assistant" && $1<"2026-10" {m=substr($1,1,7); usd[m]+=$14; n[m]++} END {for (m in usd) printf "%s  $%.2f  %d turns\n", m, usd[m], n[m]}' claude.csv | sort
2026-07  $6056.39  39639 turns
2026-08  $5898.03  28001 turns
2026-09  $2093.18  9071 turns

The output above is the original 1 October reading, before the September duplicate was removed. The corrected totals are in the chart above.

Your months will not look like mine. Install Caprock and read your own; it starts with the sessions Claude Code still has on disk.