Is Claude Code worth paying for?

Is Claude Code worth paying for?

Claude Code is worth paying for when it gives back more of your hours each month than the plan costs at your own hourly rate. That is the whole test. Every other version of this question turns into an argument about taste, and this one has an answer you can check against your own calendar.

The arithmetic is small. The judgement is in what you count as a recovered hour, and in the cases where the agent gives back nothing at all. Both are below. For the current price of each tier, see what Claude Code costs on each plan.

How many hours does a Claude Code plan have to save?

Divide the plan price by what one hour of your time is worth. The result is the number of hours the agent has to give back each month before you are ahead. As of August 2026 the paid plans are $20 a month for Pro, $100 a month for Max 5x and $200 a month for Max 20x, and all three include Claude Code. The free plan does not.

[
  {
    "label": "$20 per hour",
    "hours_for_pro_20": 1,
    "hours_for_max_100": 5,
    "hours_for_max_200": 10
  },
  {
    "label": "$50 per hour",
    "hours_for_pro_20": 0.4,
    "hours_for_max_100": 2,
    "hours_for_max_200": 4
  },
  {
    "label": "$90 per hour",
    "hours_for_pro_20": 0.2,
    "hours_for_max_100": 1.1,
    "hours_for_max_200": 2.2
  },
  {
    "label": "$150 per hour",
    "hours_for_pro_20": 0.1,
    "hours_for_max_100": 0.7,
    "hours_for_max_200": 1.3
  }
]

Read the first row. At $20 per hour, the $20 plan has to save you 1 hour in a month and the $200 plan has to save 10. At $50 per hour the $200 plan needs 4 hours. At $150 per hour it needs 1.3, which is less than one morning of work in a whole month. The chart shows 4 rates. Put your own rate in and the shape does not change: the cheap plan clears its cost almost at once, and even the expensive one is a low bar for anyone who writes code for a living.

If you are salaried and do not know your rate, divide what you cost your employer in a year by about 2,000. If you code on your own time, use what an evening back is worth to you. A rough figure is enough, because the answer is not close.

What actually counts as a recovered hour

Review time comes out of the saving. An hour of generated code that takes you forty minutes to read carefully saved you twenty minutes. Count it that way, or the whole exercise is decoration.

Work where the agent reliably gives hours back:

• Code that copies a pattern already in the repository: a new endpoint shaped like the four that already exist, or another CLI (command line interface) flag.
• Mechanical changes across many files, where the change is obvious and the tedium is the whole cost.
• Tests for code that already has tests, because the existing tests show the shape and the assertions to make.
• Reading unfamiliar code to answer a question, such as where a setting is applied, which is search rather than authorship.
• One-off scripts you would otherwise skip writing, so the saving is a task you never did by hand at all.

Those five have one thing in common. The agent is fast when correctness is cheap to check: a failing test, or a diff against a pattern already in the repository. That is the mechanism, and it is worth stating plainly because it predicts the failures. Where checking is expensive, the saving shrinks toward zero no matter how good the generated code is.

Where a Claude Code plan does not pay

An unfamiliar codebase with no tests. You cannot tell a correct diff from a plausible one, so you have to verify by reading everything the change touches. That takes about as long as writing it yourself, and you still carry the risk that you missed something. Add tests first, then bring the agent in.

Work that is mostly design. If the hard part is deciding what to build, the typing was never the bottleneck, so removing the typing saves very little. The agent is fast at implementing a decision you have already made.

Anyone who will not review the diffs. Unreviewed generated code does not save time. The cost arrives in a later week, when somebody has to work out why a function nobody read behaves the way it does. If you know you will not read them, no plan is worth buying.

Very light use. If you reach for the agent a few times a month, per-token API (application programming interface) billing costs less than any subscription. The comparison between API billing and a subscription is worth doing before you commit to a monthly charge.

Why people say "do not buy the Max subscription"

The complaint is common enough to answer directly, and part of it is correct.

The strongest point is about limits. Paid plans have a rolling session window plus weekly caps, and Anthropic has changed those limits more than once: it doubled the five hour session limits in May 2026 and has run several temporary increases to the weekly caps since. What you buy is a multiple of the Pro allowance, so the same $200 can buy a different amount of work in December than it bought in August, with no change to your bill. That is a real risk and no arithmetic removes it. How the session and weekly limits work is the detail behind it.

The second point is about the jump to $200. Most people who buy the top plan do so before they have hit the limits of the cheaper one, and then find they were never going to. That plan is priced for somebody running the agent most of the working day, often in several sessions at once. If that is not your week, you are paying for headroom you will not touch. The choice between the $100 and $200 Max plans comes down to how often you actually hit the wall.

The third point is quality, and it deserves to be taken seriously. The agent will produce a confident, well formatted, wrong change, and it will not flag which one this is. Put review time into the arithmetic above and the point is handled. It is the main reason a saving is smaller than it first looks.

What the complaint gets wrong is the opening move. Almost nobody has to choose between $100 and $200 on day one. Start at $20, use it for a month, and upgrade when you hit limits during work you needed to finish.

What changes when the agent runs on a VPS

Run Claude Code on a small server instead of your laptop and one thing changes that matters for this calculation. The work continues when you close the lid, because the session lives in a terminal multiplexer on the server and your laptop is only a window onto it.

tmux new -s claude

Start the agent inside that session, then press Ctrl-b followed by d to detach. The session keeps running on the server. Come back from any machine and reattach:

tmux attach -t claude

If your SSH (secure shell) connection drops on a train, nothing is lost, because the process was never tied to your connection. A long refactor that would have died with your laptop's sleep timer finishes while you are somewhere else. Running Claude Code in tmux on a VPS covers the setup, and the wider case for running a coding agent on a server covers why you would want more than one session at a time.

Two honest limits. A VPS does not make the model faster, because the work happens in Anthropic's data centre and your server is only running a terminal client. A VPS also does not raise your usage limits, since those are tied to your account rather than to the machine you type on.

The rent counts too. As of August 2026 a small VPS with 2 vCPU and 4 GB of RAM costs roughly $5 to $15 a month, so it belongs in the same sum as the subscription. On the $20 plan, $10 of server rent raises your break-even by half. On the $200 plan it barely moves the number. What a VPS actually costs per month has the ranges.

One more reason to use a separate box: an agent with shell access runs what it is told to run. Keeping it away from anything you cannot rebuild is cheap insurance, and the safe way to run Claude Code on a VPS describes the isolation worth setting up first.

Measure your own break-even in one month

Do not settle this by reading opinions, including this one. Buy the $20 plan and collect the evidence.

1. Keep a plain text file, one line per task, with two numbers: the minutes the task would have taken you by hand, and the minutes you spent reviewing the result.
2. Run /usage inside Claude Code to see the current session's token usage and cost. /cost is an alias for the same command.
3. At the end of the month, total the minutes you saved and subtract the minutes you spent reviewing.
4. Compare that total against the row in the chart above that matches your rate.

If you cleared the bar in a month, the question is closed and the plan is worth it. If you did not, find out why before upgrading, because a bigger plan does not fix a codebase the agent cannot check itself against. Where your tokens actually go helps if the usage figures look higher than the work you got.

The verdict for each kind of developer

You bill by the hour. Buy it. At any realistic rate the $20 plan clears its cost in one afternoon, and it is the easiest yes on this page.

You are salaried, in a codebase you know, with tests. Buy the $20 plan. Move to $100 after you hit the session limit twice in a week during real work, and not before.

You run the agent most of the working day, often in parallel sessions. The $200 plan fits, and this is the one case where it clearly does. Price the API route anyway, because heavy but bursty use sometimes lands cheaper per token.

You are a student or a hobbyist coding on your own time. The $20 plan, or the free API tier for occasional use. $100 a month is hard to justify against an evening rate.

Your day is mostly design and review. Probably not at $100. The bottleneck in that week is decisions, and the agent does not make those. The $20 plan still earns its keep on the small tasks around the edges.

You will not read the generated diffs. No plan is worth buying. Fix that first.

If you are still choosing between products rather than plans, the comparison with Cursor, Codex and Copilot is the next thing to read, and the plan by plan breakdown covers the tiers in more detail.

FAQ

Is Claude Code included with a Claude Pro subscription?

Yes. As of August 2026, Claude Code is included with Pro at $20 a month and with both Max tiers at $100 and $200 a month. The free plan does not include it. Pro is also the right place to start, because the break-even at $20 is roughly one recovered hour a month for most people, and one month of use shows you what your real usage looks like before you commit more.

Should I start on the $100 or the $200 Max plan?

Neither. Start on Pro at $20 and use it for a month of ordinary work. Upgrade when you hit the usage limits repeatedly during tasks you needed to finish. Buying the top plan first means paying for headroom before you know whether your week has any use for it.

Why do people say the Max subscription is not worth buying?

Two of the reasons hold up. The plans sell a multiple of the Pro allowance rather than a fixed amount of work, so what your money buys shifts when the underlying limits shift, and they shifted more than once during 2026. And most buyers of the $200 plan never reach the limits of the $100 one. The other common complaint, that the agent writes confident wrong code, is true, but it argues for counting review time rather than for skipping the plan.

Does running Claude Code on a VPS give me more usage?

No. Usage limits are tied to your Anthropic account, so the same limits apply whether the agent runs on a laptop or on a server. What the server buys is continuity: the session survives a closed lid and a dropped connection. The rental cost, commonly $5 to $15 a month for a small instance as of August 2026, belongs in your break-even sum alongside the subscription.