My Claude bill is $200 a month. It was $200 last month and it will be $200 next month, whether I run one session or five agents across three worktrees. Kentucky adds 6%, so I pay $212.
That number has never once been the reason I changed plans. I moved up because the plan below kept stopping me in the middle of things.
What every comparison gets structurally wrong
Search claude code pro vs max and you get arithmetic. Messages per five-hour window, cost per
unit of Pro-equivalent usage, break-even tables. ClickRank,
Layer3Labs and
Serenities AI all run some version of it.
The arithmetic is not wrong. It is answering a question the product does not pose. A capped plan has no cost curve. There is no month where I use it harder and pay more, so there is no calculation whose output is “upgrade”. The only thing that varies is whether I get stopped.
What the guides compare: dollars per message, dollars per Pro-equivalent unit, break-even volume
What actually differs: how far into a working day you get before the window closes
What getting stopped looked like
On the plan below, I hit the limit and the cap regularly. Not at the edge of some heroic session. On ordinary days, doing ordinary work.
The part that made it expensive was never the waiting. It was that the stop lands mid-task. An agent is four files into a change, or a worktree is halfway through a gate, and the next request does not go. You come back later and rebuild the context you had.
Three mechanics make this worse than it sounds, and all three are in Anthropic’s own documentation rather than something I inferred:
| Mechanic | What it means | Source |
|---|---|---|
| A rolling 5-hour window plus a weekly one | Two ceilings, and the weekly one is the surprise | Pro and Max plan limits |
| Shared across every Claude surface | Chat on claude.ai spends the same allowance as your terminal | Usage and length limits |
| Switching model does not restore access | The windows are shared across models, so /model is not an escape hatch | Manage costs |
That third one is the one I would want to have known. The instinct when you get stopped is to drop to a cheaper model and keep going. It does not work, because the ceiling is not per-model.
Why I never ran the numbers
I could not have, and neither can you.
There is no per-token record of subscription usage to look at. /usage reports the last 24 hours
or the last 7 days, computed from local session history on the machine you run it on, so it cannot
produce a month and it does not see your other machines. Anthropic
says so directly: the session cost figure “isn’t relevant
for billing purposes” for Max and Pro subscribers.
If you are on the API instead, where the meter is real, the arithmetic does exist and I have built it: a cost calculator that prices a job from units of work rather than token counts, using counts measured on a real 2,399-page run.
So the honest version of “should you upgrade” contains no calculation at all:
The rule I actually used
Do
- Upgrade when the limit stops you mid-task, repeatedly, on ordinary days.
- Try the middle tier first. Most people who think they need 20x need 5x.
- If you are never stopped, you are buying headroom you will not use, and the guides are right about that.
Don't
- Upgrade because a table says your volume justifies it. The bill is flat; volume justifies nothing.
What the number is worth against a benchmark
Anthropic publishes its own figure, which is the most useful calibration I have found: $13 per developer per active day, and $150 to $250 per developer per month across enterprise deployments, with 90% of users under $30 per active day (Manage costs).
A capped $200 sits at the top of that band. That is the whole product. You are not buying cheaper tokens, you are buying a ceiling on the invoice, and the trade is that there is a different ceiling on the work.
Where cost genuinely is the question
Not here. The moment cost becomes real arithmetic is when work moves off the subscription and onto the API, which is what happens as soon as you point a crawler, a batch job or a fleet of workers at a model. That has a per-token bill, it varies enormously with how you shape the job, and I measured one: the same audit cost $452 and $47 depending on whether it ate raw HTML or extracted text.
Two habits from my own setup push usage up hard enough to matter on either meter. Running several agents on worktrees multiplies consumption by roughly the number of agents, and Anthropic notes agent teams run about 7x a standard session when teammates work in plan mode. And a second model reviewing the first is a second bill, or a second draw on the same window.
What I cannot tell you
I also cannot give you my usage in tokens, for the reason above, and I would rather say that than publish an estimate shaped like a measurement. If you want a number to plan against, use Anthropic’s $150 to $250 band and your own answer to one question: how often does the thing stop you?
Sources
- Use Claude Code with your Pro or Max plan, Anthropic Help Center
- Models, usage, and limits in Claude Code, Anthropic Help Center
- Extra usage for paid Claude plans, Anthropic Help Center
- Manage costs effectively, Claude Code docs
- Plan pricing, read 2026-08-21
Related:
- Claude Code API Cost: $452 vs $47 for the Same Job: the other meter, where the arithmetic is real
- Claude Code Worktrees: 4 Gaps in the Default Setup: the practice that makes a ceiling arrive sooner
- Fable as the Driver: what a second model actually costs to consult
- Claude Code on a Raspberry Pi 5: Setup and 3 Traps: the same flat plan, running headless on a $173 box with no API key