The question about plans usually turns up in week two, not week one. First you set up an account, run a few conversations, everything works. Then a limit message lands in the middle of something, and the pricing page stops being an abstraction.
Except pricing pages compare prices, while the decision actually rests on something else entirely – the rhythm of your own work. What follows is an attempt to translate one into the other.
There isn’t one limit, there are two
Claude counts usage in two windows at once: a rolling five-hour one and a weekly one. The first refills a few hours after heavy work begins, the second once a week. Both are shared between browser conversations and Claude Code, so an hour spent in chat draws on the same pool as an hour in the terminal.
A shared pool doesn’t mean both routes are treated the same, though, and that’s an argument for the terminal you rarely hear. In May 2026 Anthropic doubled the five-hour limit specifically for Claude Code, and dropped the restriction that used to trim Pro and Max accounts inside Claude Code at busy hours. Both changes applied to the terminal only and left browser conversations untouched. Anthropic announced them as permanent rather than as a temporary promotion – and that’s how we read it: for the same bill, work done in the terminal fits inside a plan more comfortably today than the same work pasted into a chat window.
The weekly limit isn’t the sum of every five-hour pool across the week – it’s smaller than that. A few days of genuinely heavy work can exhaust it before the week is out.
Pro sets the reference level, and Max is sold in exactly that currency. It comes in two tiers, and the names say it outright: Max 5x and Max 20x – five times or twenty times more usage than Pro. Not faster, not smarter: longer, before the tool says enough.
And one thing that settles the question straight away: Claude Code isn’t available on the free plan. If what you were really asking was whether to stay on free, and you want to work on your own files, the answer is binary – Pro is the entry point.
Anthropic doesn’t say how much that is in tokens
What you’ll find in the plan materials is one tier’s multiplier against another, plus a line saying paid plans get higher limits. No actual number appears, for Pro or for Max. That’s disorienting when you’re trying to budget, but there’s a simple logic behind it: a subscription works a bit like a mobile phone contract. The pay-as-you-go card here is paying for Claude through the API – you settle up token by token and know to the penny what every minute cost you. On a contract you buy a bundle and stop counting. For everyday work it’s the card that comes out more expensive: Anthropic’s own documentation puts per-person monthly figures for token billing above what Max costs.
The analogy breaks in one place: you cancel a Claude subscription from one month to the next, with no two-year contract and no notice period counted from the day you signed.
What else separates Pro from Max
There’s a difference that applies from the first minute and is easy to miss: your default model depends on your plan. On Pro a session starts on Sonnet, on Max on Opus. You switch with /model in the terminal, and the pool is shared across every model – except that a heavier model eats through it faster. If you’re starting on Max, though, we’d suggest moving to Sonnet for the first few weeks: it’s faster and perfectly sufficient for everyday work. Keep Opus for analysing large sets of data and for multi-hour sessions on one task – we went through that in the post on choosing a model.
Max also buys early access to new Claude features – the pricing page calls it exactly that. From what we’ve seen, a feature that shows up on Max first becomes available on Pro later on. So if a single new toy is the only thing pulling you towards the higher tier, waiting is sometimes the cheaper move.
First pattern: a few short sessions a day
In the morning you have Claude Code go through a spreadsheet export and work out how many hours went into which project last month. After lunch you come back with an error from a script that worked yesterday, and the two of you spend a quarter of an hour finding out why. In the evening you have it fix a repeated passage across three files on your site and check whether it appears anywhere else. Each of these takes anywhere from fifteen minutes to an hour and sits inside its own window – in between you’re doing something else entirely.
At that rhythm Pro is plenty. The five-hour window has time to refill before you come back, and you won’t get halfway through the weekly one. If that’s what your day looks like, there’s probably no point paying for a higher tier.
Second pattern: always enough, but only just
It also happens that the limit never actually runs out, but all week you have exactly as much of it as you need. Nothing hurts, you just glance at the bars now and then and work out in your head whether it’ll last until Friday. This is the hardest one to settle, because nothing in it shouts.
You can play the model here, and plenty of people do. /usage shows not only how much of the window is left but when it refills – and those two numbers together give you a hint. When there’s plenty left and the window refills in half an hour, you move to Opus and hand it the hardest task on the list, because the rest of those tokens will go to waste anyway. When there’s little left and the refill is a long way off, you drop to Haiku – the lightest of the models – and use it for the simple, repetitive work.
Switching models invalidates the prompt cache – each model has its own, so right after the change the next request reads the whole conversation from scratch, even though not a word of it moved. Before you switch models mid-session, run /compact: Claude Code summarises the conversation so far and you carry on with a much shorter context. Or simply start a new chat, if the next task doesn’t need what came before it anyway.
Balancing works, but it has a price. If it’s turning into a daily ritual, or you catch yourself holding back your own work to avoid burning the limit, that’s when Max is worth considering.
Third pattern: half a day inside one Claude Code session
You open Claude Code in the morning, work on one project, the session stays open for four hours, and Claude Code reads files, runs commands, fixes things, checks them.
On Pro this kind of work finds the ceiling quickly. Halfway through the day you stop and wait for the five-hour window to refill. The weekly limit runs out a few days too early. You find yourself paying for extra access at API rates to get a task finished. If that’s what your week looks like, this is the moment for Max.
Before you upgrade, though, check /usage – it shows your plan usage bars and a breakdown of what’s eating them, and d and w switch the view between the last day and the last week. Sometimes the culprit isn’t the amount of work but one session dragged along since Monday, carrying a week’s worth of conversation into every question.
Settle the choice between Max 5x and 20x the same way, not on a hunch. Two weeks on 5x will tell you whether you touch that ceiling at all.
Where to begin, and when to change
Start on Pro. In the first few weeks you won’t gather enough data to judge whether you need a higher tier. The signal to consider changing is a simple one: you hit the limit regularly, not once, in an unusually busy week.
It’s worth keeping in mind that moving up to Max isn’t an irreversible step. On monthly billing the plan moves both ways: ahead of a month you know will be heavy – a big migration, a site move, a project with a deadline – you take Max, and when the pace returns to normal you drop back to Pro. Nobody asks you to declare a year in advance how you’re going to work.
There’s also a middle option: usage credits, which let you keep working past your plan’s limit, billed separately – you turn them on in the terminal with /usage-credits. Sensible when you hit the limit occasionally and don’t want to jump a tier for two days a month. One caveat: on credits the prompt cache lifetime drops from an hour to five minutes, so longer breaks inside a session get pricier than usual.
Before you click annual billing
Annual is cheaper – paying a year up front takes the monthly Pro rate down by a percentage in the teens. Not a staggering saving, but a saving.
One note from our own practice at JUSTAUTOMATE before you click it. We changed how we work with Claude twice in the past six months, and each time it changed the answer to “how much do we actually need”. A year is a long time at that pace. If you’re just starting, paying monthly for the first couple of months costs next to nothing extra and buys something worth more – the right to change your mind.
In short
Start on Pro and give yourself the first month to find out how you actually work – you don’t know that on day one. If it turns out you hit the limit regularly, stop halfway through the day, or pay for extra access at API rates to finish a task, take Max rather than working around it. If you only brush against the ceiling now and then, waiting for the window to refill is enough, or buying credits when a deadline is closing in. And if none of that describes you, stay on Pro and come back to the decision when the way you work changes. The plan moves either way from one month to the next, so there’s no need to buy the expensive one at the very start of your time with Claude.