Claude Max subscribers are auditing Anthropic’s meter

Claude Max subscribers who count their own tokens in Claude Code’s local logs report getting less usage than they paid for, and a September 28 write-up on Nooneshappy gathers their GitHub reports along with at least three class actions over the plans. Because Anthropic publishes no token figure for Max 5x ($100 a month) or Max 20x ($200), those logs are the one independent check on the meter, and Claude Code is the only Claude product that keeps them: chats on claude.ai and the mobile app leave nothing to count, and none of the reports has been independently verified.

Counting from the logs

Claude Code, including the Code tab in the desktop app, saves every session as a JSONL file under ~/.claude/projects/, and each response records message.usage with input, output, cache-write, cache-read and thinking tokens. One response is written to the log as several lines with the same figures, so totals have to be deduplicated by requestId. Skipping that roughly doubles the count, a mistake at least one poster made and corrected in public. The meter’s own reading is stored as cachedUsageUtilization in ~/.claude.json.

Dividing the tokens used by the points the meter moved gives the price of 1% of the limit: if 5.9 million tokens take the weekly gauge from 0% to 7%, 1% is about 842,000 tokens. Some posters also weight token types by Anthropic’s API prices, with input at 1x, cache writes at 1.25x, cache reads at 0.1x and output at 5x; the write-up calls that a guess, since Anthropic doesn’t publish its weighting.

The write-up’s simplest option is ccusage, an open-source tool that reads the same logs: npx ccusage@latest weekly totals tokens by week, and ccusage blocks groups them into Claude Code’s five-hour windows. Its suggested check is to note the /usage percentage right after a weekly reset and again a few days later, then set the change against ccusage’s total for the same days. The logs cover only the machine they sit on, so usage from another computer, claude.ai or the mobile app stays out unless added by hand.

What does 5x buy?

One subscriber who moved from Pro to Max 5x found that 1% of the week cost 612,000 tokens on Pro and about 840,000 after the upgrade, against the roughly 3 million expected on Max 5x. Each point on the weekly meter was worth only about 1.4 times what it had been on Pro, and the app still says “Pro”.

That 3 million expectation applies the plan’s fivefold multiple to the weekly meter. The class action filed September 8 alleges that the 5x and 20x figures cover only the five-hour session window, not the weekly limit subscribers actually run into. Anthropic’s help article on Max usage words the tiers as “5x more usage per session than Pro” and “20x more usage per session than Pro”, with the session limit resetting every five hours, and says Anthropic “may limit your usage in other ways, such as weekly and monthly caps or model and feature usage, at our discretion.”

The 20x tier draws the same complaint. A July report has a Max 20x weekly limit draining at the 5x rate or faster after an upgrade, and a subscriber who went from 5x to 20x used 1.3 times more yet ran near the weekly limit more often. A team that pools Claude Code subscriptions, using its developer’s own normalization, put the weekly gap between the two tiers at about 1.5x in September, down from about 2.5x in late August.

On Max with Opus 5, one report found the same work used 1.8 times more of the five-hour limit when run from a script with claude -p than in the app. On Max 20x, another says Fable 5.1 produced twice the output per turn of Fable 5, and the weekly meter went from 0% to 88% in 22 hours.

Kahn v. Anthropic, filed in June, alleges more broadly that “the actual usage provided by the Max 5x and Max 20x plans is far below the advertised amount of usage.” Pascual v. Anthropic alleges that backend changes between March and May 2026 reduced the value of Pro and Max; its complaint, filed July 24 in the Northern District of California, says subscribers couldn’t figure out whether drops in their remaining usage percentage came from legitimate requests or a defect, Bloomberg Law reported.

The September 11 jump

A Max 5x subscriber running Claude Code on two machines, with Opus 5 at the xhigh effort setting for about two months, logged a break between 01:30 and 03:30 UTC on September 11. From then on, every active session on both machines put a thinking block on 95% to 100% of requests, against 30% to 67% before, and average output per request, 250 to 960 tokens before, rose to 2,500 to 3,000 tokens in the first hours. Both machines ran client 2.1.260, the model, effort, sessions and kind of work were the same, and no settings, hooks or project instructions had been edited in the 19 hours before. The figures come from Claude Code’s transcripts, deduplicated by request ID with subagents excluded.

A full five-hour window normally took 40% to 50% of their session limit; after the change, 24% went in 36 minutes, a pace that would hit the limit in about two and a half hours. The subscriber usually reaches the weekly limit, which resets on Monday, with about 2% to spare, and was at 97% by Friday evening with three days still to go.

A second user, calling Claude Code non-interactively from a Linux server with a fixed prompt template, logged the same break. Their last normal call, on client 2.1.263 on September 10, spent 40 thinking tokens on a 43,000-token input; from their first abnormal call, early on September 11 on the same binary, requests carried 25,000 to 35,000 thinking tokens each. Run side by side at the same minute on the same prompt, client 2.1.237 produced no thinking tokens and 25,148 output tokens in 5.4 minutes, while 2.1.268 produced 16,128 thinking and 48,586 output tokens in 10.3 minutes.

The write-up reads the overnight change on an unchanged client as pointing to Anthropic’s side, possibly rolled out in phases, and the gap between versions as a sign that the client plays a part too: a change on the server in how much the model thinks at a given effort setting could interact with whatever each client version sends. The subscriber asked Anthropic to announce changes of this kind and to offer a way to opt out.