Claude Max: decide before you upgrade

Should you upgrade to Claude Max at all?

Upgrade to Claude Max when one week of your own records shows you stopped by a usage limit, repeatedly, on work that could not wait for the reset. If your week does not show that, a larger plan buys you nothing you will notice. The input to this decision is your own usage, and a feature comparison cannot supply it.

The plans are hard to separate on features because they largely sell the same product in different amounts. So the useful question is about you: how often does the plan you already pay for stop you, at what hour, on which kind of work, and how much of your ceiling goes on work a cheaper path would have handled?

Below is a procedure you can run in one week, and a map from each result to an answer. Two of the answers are not Max.

Log four things every time Claude stops you

Run the week without changing your habits. If you change how you work and measure at the same time, you get a week you cannot read.

Every time Claude refuses to continue, write one line. Four fields are enough.

• The exact limit message, because the wording decides what you can do next.
• The local hour, because a block at 22:00 on a Sunday is a different problem from one at 10:00 on a Tuesday.
• Whether the blocked work could have waited for the reset. Answer yes or no, honestly.
• The model you were on when it happened.

Put a small function in your shell so recording a block takes five seconds. A block you did not write down did not happen.

# ~/.bashrc
climit() {
  printf '%s\t%s\t%s\t%s\t%s\n' \
    "$(date +%Y-%m-%dT%H:%M)" "$1" "$2" "$3" "$4" \
    >> "$HOME/claude-limit-log.tsv"
}

Reload with source ~/.bashrc, then record a block in one command. The argument order is limit, model, could-it-wait, description.

climit weekly opus no "payments test migration, due Thursday"

At the end of the week, read the file back.

cut -f1 "$HOME/claude-limit-log.tsv" | cut -dT -f2 | cut -d: -f1 | sort | uniq -c | sort -rn
cut -f2 "$HOME/claude-limit-log.tsv" | sort | uniq -c | sort -rn
awk -F'\t' '$4 == "yes"' "$HOME/claude-limit-log.tsv" | wc -l

The first line counts blocks by hour of the day. The second counts them by limit type. The third counts the blocks that landed on work which could have waited. If that third number is most of your week, Max would buy you speed you did not need. The fix is scheduling, and scheduling is free.

What does the limit message actually tell you?

Claude Code prints one of four messages when you run out of allowance.

You've hit your session limit
You've hit your weekly limit
You've hit your Opus limit
You've hit your Sonnet limit

The session and weekly messages describe a plan allowance shared across all models, so switching with /model will not restore access. The Opus and Sonnet messages are specific to one model family, so switching to a model outside that family with /model does keep you working. That distinction is documented in the Claude Code cost guide at code.claude.com/docs/en/costs, read on 29 August 2026.

That difference is the whole reason to write the exact wording down. A week full of session limits with no weekly limit is a pacing problem, and pacing is free to fix. A week that ends against the weekly limit is a volume problem, and volume is the thing a larger plan actually sells.

One more fact changes how you read your own log. Anthropic's help centre states that "Pro and Max plans offer usage limits that are shared across Claude and Claude Code, meaning all activity in both tools counts against the same usage limits." So a chat session on claude.ai and a coding session in the terminal draw on the same ceiling. If you have been blaming Claude Code for your limits, check your chat use before you pay to raise them. how the session and weekly windows fit together covers the shape of the windows themselves.

How much of your ceiling goes on work a cheaper path would do?

Run /usage inside Claude Code. On a Pro or Max plan it shows your plan usage bars along with a breakdown of what consumed them. Press d or w to switch between the last 24 hours and the last 7 days.

Read the caveat before you read the numbers. The documentation says these figures are approximate and computed from local session history on this machine, so usage from other devices and from claude.ai is not included. Your handwritten log records when you were blocked. /usage records where the tokens went on one machine. You need both.

Two parts of that breakdown do most of the explaining.

Attribution splits recent usage across skills, subagents, plugins and individual MCP servers, each as a share of the total. An MCP server's share counts only the requests that consumed one of its tool results, so a large share means you are genuinely using it, and /mcp lets you disable the servers you are not.

Behaviour flags appear when a single behaviour accounts for 10% or more of recent usage. Long context and cache misses are the two you are most likely to see.

Long context has a plain cause. Claude Code sends the whole conversation with every request, so a one-line question in a session you opened this morning still draws usage for the entire history behind it. /clear between unrelated tasks resets that and costs nothing. /compact is a poor substitute, because it reads the conversation it summarises, which makes compacting a large context a large request in itself.

Cache misses have a plain cause too. Your first message after a break longer than the prompt cache lifetime reprocesses your full context instead of reading it back from cache. The documentation puts that lifetime at one hour on a subscription, dropping to five minutes once you are drawing on usage credits.

If a large share of your week sits in long context, or in Opus left as the default model, you have found ceiling that a cheaper path would absorb for free. what a Claude Code session actually spends tokens on goes through the rest of the breakdown, and matching the model to the job is usually the largest single change available to you.

What does Claude Max cost if you pay from France?

Read the checkout total, then read the first invoice. Any euro figure printed in a guide is wrong within a month, because two of the numbers on your bank statement are not set by Anthropic.

As of 29 August 2026 the plan cards on claude.com/pricing are shown in US dollars, under the footnote "Prices shown don't include applicable tax." Pro is listed at $20 per month billed monthly, or $17 per month billed annually with $200 charged up front. Max is listed as "From $100" per month, and it has two tiers, named Max 5x and Max 20x in the usage credits documentation. the difference between the two Max tiers is a separate decision, and it only becomes relevant once your log has already said yes to Max.

The first number that is not on the plan card is tax. Anthropic's help centre says tax is calculated from your billing address, and that the billing address is taken from your payment method's address. So the card you pay with decides your tax jurisdiction, not the country you are sitting in. If you hold a valid VAT number, an optional Tax or VAT ID field appears during signup when your address is eligible for one, and afterwards under Settings, then Billing, then the Update button next to your payment method.

The second is your bank. A euro card paying a dollar amount converts at the issuer's rate, and many cards add a fee for a transaction in another currency. That line comes from your bank's own fee schedule and never appears on the Anthropic invoice. Compare your bank statement against the invoice for the same month once, and you will know your real multiplier for every month after.

The invoice lives under Settings, then Billing, then the Invoices section, with a View button per invoice, and a copy is emailed to your billing email address. The lines worth reading are the plan charge, the tax line, "Applied balance" for any account credit spent on this invoice, and "Amount due", which is the total minus that applied balance. A mid-cycle upgrade adds a "Prorated charges" line, described as one full billing cycle of the new plan less a prorated amount, plus a "Credits" line holding the leftover value from the old plan. Usage credits are billed separately and get their own receipts.

What does leaving cost, before you pay to enter?

Settle this before you upgrade, because it sets how expensive a wrong guess is.

Cancelling takes effect at the end of the current billing period, and you keep the paid plan until then. To avoid the next charge, cancel at least 24 hours before the next billing date. Leftover value from a plan change comes back as account credit applied to a later invoice, which is why "Applied balance" exists as a line item. Your card is not refunded.

Refunds are narrower. Anthropic's refund article states that "except as expressly provided in our Consumer Terms of Service or where required by law, all payments are non-refundable", and that customers in the EEA and the UK, which includes France, "may receive a refund within 14 days of purchase", prorated according to use. Asking for a refund does not cancel the subscription, so someone who wants out has to do both. All of this was read on 29 August 2026, and billing policy is exactly the kind of thing that changes, so confirm it in your own Settings before you rely on it. the step by step for cancelling or moving back down a plan carries the current procedure.

Two consequences for your test. Buy Max monthly rather than annually, because an annual plan charged up front commits you for a year against a decision you are still testing. And if you subscribed through the iOS App Store, Anthropic support cannot refund that charge at all, because only Apple can.

Map your week to an answer

Read the row that matches your log.

• No limit message all week. Stay where you are. You are paying for a ceiling you never touch, and the honest move is to check whether you need Pro either. what Pro includes and where its limits sit is the comparison to make first.
• Session limits only, clustered into one or two bursts, with most blocks marked "could have waited". Stay on Pro and change how you work. Clear between unrelated tasks, and move the work that does not need Opus onto a smaller model. If your hour counts show one clear peak, shift the heavy run out of it. Then run the same log for a second week and see whether the blocks survive the change.
• The weekly limit, more than once, on work with a date attached to it. This is the case Max exists for. The ceiling is volume, your log proves you need the volume, and no amount of pacing gets it back.
• Limits hit hard but rarely, for example one deadline week per month. Usage credits are usually cheaper than twelve months of Max bought for two busy weeks.

Usage credits let a subscriber on Pro, Max 5x or Max 20x keep working past the included limit. They are prepaid, billed at standard API rates, and charged separately from the subscription. Turn them on under Settings, then Usage on claude.ai, or run /usage-credits in Claude Code, which opens the same page. Set a monthly spend cap instead of choosing unlimited, because the reason you are here is that you want a ceiling you picked yourself. One caveat is worth knowing before you switch: while you draw on usage credits the prompt cache lifetime drops from one hour to five minutes, so a day of short scattered sessions costs more per question than the same day inside the plan. how metered API spend compares with a flat subscription is the arithmetic behind that choice.

Whatever the log says, keep it running through your first month on the new arrangement. The same four fields tell you whether the change fixed the thing you bought it for, and that is the only evidence that a renewal is worth signing.

FAQ

How long should I log before deciding on Claude Max?

One full week, and it has to include your busiest day. A shorter sample describes one project rather than your work. If the week was unusually quiet or unusually heavy, run a second one before deciding. An extra week of measuring costs nothing, and a wrong upgrade costs a billing period.

Does /usage in Claude Code show my whole account?

No. The documentation states the figures are approximate and computed from local session history on that machine, so a second laptop and everything you did on claude.ai are missing from it. For account-level numbers, open Settings, then Usage on claude.ai. Use /usage for the breakdown of where tokens went, and your own log for when you were blocked.

Will switching model get me past a session or weekly limit?

No. The session and weekly allowances are shared across all models, so /model changes nothing while one of those is exhausted. The model-specific messages behave differently: after "You've hit your Opus limit", switching to a model outside that family with /model does let you keep working.

Is it cheaper to add usage credits than to upgrade to Max?

It depends on how many months look like your worst week. Usage credits are prepaid and billed at standard API rates, so they cost nothing in a quiet month, while a subscription charges every month whether you reach the ceiling or not. Credits also cost more per question during a burst, because the prompt cache lifetime drops to five minutes while you draw on them.

I am in France. Which figure should I budget from?

The total on the checkout screen, and then the first invoice under Settings, then Billing. Checked on 29 August 2026, the plan cards on claude.com/pricing were listed in US dollars with a note that tax is not included, tax is calculated from the address on your payment method, and your card issuer adds its own conversion rate and possible non-euro transaction fee on top. Only the invoice and your bank statement together give you the real monthly number.