Most guides to Claude’s pricing are a table of tiers and prices, which is the least useful part. All the paid plans give you the same product — Chat, Cowork, and Claude Code. What differs is how much you can use before something stops.
So the useful version of this guide isn’t “which tier?” It’s “how do the limits actually work, and what’s eating mine?”
The plans, briefly
At the time of writing:
| Plan | Roughly | Who it’s for |
|---|---|---|
| Free | £0 | Chat only — no Cowork |
| Pro | Individuals. Full product, tightest limits | |
| Max 5x | ~$100/month | Daily heavy use; the realistic floor for full-time work |
| Max 20x | ~$200/month | All-day agentic work, long-running tasks |
| Team | ~$25/seat/month | Organisations — commercial data terms, admin controls |
| Enterprise | Custom | SSO, audit logs, role scoping, compliance options |
Two things matter more than the numbers.
The free plan doesn’t include Cowork. If you’re evaluating whether Claude can do genuine multi-step work on your files — which is the thing most people are actually trying to assess — the free plan can’t show you. You’ll judge a chat window and conclude something about a product you never saw. What Cowork is.
Team and Enterprise aren’t just bigger Pro. The jump from Max to Team isn’t primarily about capacity — it’s about commercial terms. Under commercial terms Anthropic doesn’t train on your prompts or code; on consumer plans it can, when data sharing is on. If this is company work, that distinction is the reason to be on a company plan, independent of usage. Details here.
How limits actually work
This is the part worth internalising, because it explains nearly every surprise.
There are two layers:
- A rolling five-hour window. Governs bursts. Use Claude intensely for an hour and you’re drawing against this. It clears as the window rolls forward.
- A weekly cap. Governs sustained use. It resets seven days from your first message, at 00:00 UTC.
Hitting one is not the same as hitting the other. A five-hour limit is an annoyance that resolves over lunch. A weekly cap on a Wednesday is a genuine problem, and it’s the one that makes people upgrade.
Capacity has been moving in users’ favour: five-hour limits were doubled in May 2026 with peak-hour reductions removed, and weekly limits have been running meaningfully above baseline for Pro, Max, and Team. Treat any specific number you read — including ones on this page — as a snapshot. The structure is stable; the quantities aren’t.
The thing that catches everyone
The allowance is shared across everything.
Claude Code, claude.ai chat, and Cowork all draw from the same bucket. A long morning of chat reduces what’s available for an afternoon of Claude Code. A Cowork task grinding through a folder of files is spending the same allowance you were saving for something else.
This is the single most common source of “why am I limited, I’ve barely used it?” — you’ve barely used that surface. You’ve used the plan.
It also means the mental model people carry from other tools is wrong. You aren’t buying a Claude Code quota and a chat quota. You’re buying capacity, and choosing how to spend it.
Find out what’s actually eating it
Rather than guessing, ask. The /usage command breaks down what’s driving your limits — by skill, subagent, plugin, and MCP server — and there’s per-agent and per-task cost attribution for finer-grained work.
Run it after a session that felt expensive. The results are frequently surprising, and the surprises cluster in the same few places:
- A background subagent doing far more work than the task warranted. Subagents run in the background by default now, which is excellent for throughput and means the expensive one isn’t in front of you.
- An MCP server returning enormous results that quietly fill your context on every call.
- A plugin or skill loading on every session whether you needed it or not.
- Effort set higher than the task deserves —
xhighon work that a lower setting would have handled identically.
You can’t manage what you can’t see. /usage is how you see it, and five minutes with it beats an upgrade for most people.
Five habits that stretch a plan
In rough order of impact:
1. Match the model to the task. Model defaults vary by plan — Sonnet 5 is the default on Pro, Team Standard, and Enterprise seats; Opus is the default on Max and Team Premium. Sonnet 5 is genuinely strong at coding and tool use, with a native 1M-token context window, and costs far less to run. Reaching for the largest model on every task is the most common way to burn a plan for no benefit. Which model to use.
2. Manage your context. A long session carries everything before it. /compact summarises and continues; /clear starts fresh. Neither is exotic — they’re the difference between a session that stays cheap and one that gets more expensive with every turn. The difference explained.
3. Dial effort to the task. Higher effort levels mean more thinking, which means more tokens. xhigh is right for genuinely hard problems and wasteful on routine ones. It’s adjustable per task; treat it as a dial, not a setting you configure once.
4. Scope the work. “Fix the login bug in auth.js” costs a fraction of “find and fix whatever’s wrong with the app.” The second sends Claude reading everything. Specificity is a cost control as much as a quality control.
5. Configure a fallback. You can set up to three fallback models, tried in order, so a session degrades gracefully instead of stopping when your first choice is unavailable. Cheap insurance for long-running work.
Choosing, without overthinking it
- Just trying it out → Pro. Full product, and a fortnight tells you your real pattern better than any estimate.
- Hitting limits weekly → Max 5x. Occasional limits are a habits problem; weekly limits are a capacity problem.
- Running long agentic tasks daily → Max 20x. If Claude is working while you do something else, most of the day, you’ll know.
- Company work → Team, at minimum, for the commercial data terms. This is a policy decision, not a capacity one.
- Regulated, or needs SSO and audit logs → Enterprise, and talk to an account team about the options that aren’t self-serve.
The framing that helps
It’s tempting to treat a usage limit as a wall the vendor put in your way. More useful is to treat it as a budget, and notice that most people’s spending is concentrated in things they’d stop doing if they could see them.
That’s the actual argument of this guide. The people who never hit limits aren’t on more expensive plans — they’re the ones who ran /usage once, discovered that a single MCP server or an overpowered default was eating a third of their allowance, and changed one thing.
Do that before you upgrade. If you still hit the wall afterwards, you’ve genuinely outgrown the tier, and the upgrade will actually fix it.