Claude Max for coding: is it worth it? The real math

Short answer: Claude Max is worth it if you code with an agent most days and your bill would otherwise be metered. At $100 or $200 a month it is a fixed number, and the API is not. If you use Claude Code a few times a week, Pro at $20 already covers you.

That is the conclusion. The rest of this is the arithmetic behind it, run against published prices only, plus the part most posts skip: what Anthropic's 2026 authentication rules actually let you point that subscription at. This assumes you have already picked the tool; if you have not, which of Claude Code or Cursor fits how you work is the decision one step earlier.

What do you actually get for $100 or $200 a month?

Two tiers. Anthropic prices Max 5x at $100 a month and Max 20x at $200, described as five times and twenty times more usage per session than Pro. Both include Claude Code, and both share one usage pool with the Claude apps. What Anthropic does not publish is the absolute number.

Two things in that table are worth saying out loud.

The multipliers are ratios of an unpublished number. Anthropic states that Max 5x is five times Pro per session and that Max plans carry two weekly limits — one across all models, one for Sonnet models only — without publishing a figure for any of them. You cannot compute your ceiling before you buy it. You find it by hitting it.

The pool is shared. Anthropic's own guidance is that Pro and Max limits are shared across Claude and Claude Code. An afternoon spent in the desktop app is an afternoon of coding allowance spent. That is fine once you know it, and annoying if you find out at 4pm.

One more thing from that same page, because it costs people real money: if an ANTHROPIC_API_KEY environment variable is set, Claude Code prefers it over your subscription and bills at standard API rates. You can pay $200 for Max and run up an API invoice on top of it, and nothing in the terminal shouts about it.

What would the same work cost on the API?

Less than folklore suggests, and more than you would guess from the sticker price. Published API rates for Claude Opus 4.8 are $5 per million input tokens and $25 per million output, with cache reads at a tenth of the input price. Run that against a realistic Claude Code turn and $100 buys roughly twelve hundred turns.

Here are the published rates, current at time of writing:

Now a turn. Assume Claude Code sends 50,000 tokens of conversation and file context, of which 45,000 are cache reads and 5,000 are fresh, and writes 1,500 tokens back. Those volumes are our assumption, not Anthropic's — stated explicitly so you can redo the sum with your own. The prices are Anthropic's.

Twelve hundred Opus turns a month is about 55 a working day. That is a genuine day's work with an agent, and it costs what Max 5x costs. So the honest headline is the one nobody selling a subscription puts in a headline: for one person at one terminal, the flat plan and the metered API are in the same neighbourhood. They are not orders of magnitude apart.

Two things move the API side of that upward:

• Sonnet 5's $2/$10 is introductory pricing, in effect through August 31, 2026, after which it becomes $3/$15. The same turn goes from $0.034 to $0.051.
• Opus 4.7 and later, and Sonnet 5, use a newer tokenizer that Anthropic says produces roughly 30% more tokens for the same text. Better output per token, more billable tokens per file.

Where does the flat rate actually win?

Parallelism. One person at one terminal generates one stream of turns, and there the two prices are close. The moment you run four agents on four branches at once, your turn count multiplies and the subscription price does not. That is the whole economic argument for a flat rate, and it is narrower than most posts admit.

Take the same 1,200 turns a month and run four agents instead of one. You are at roughly 4,800 turns, which is about $400 of Opus at published rates against $100 or $200 flat. Nothing about the model changed. Only how many of them you kept busy. Keeping four busy means four isolated checkouts, and Claude Code will start each session in its own worktree for you.

The honest caveat is that at that volume you stop being limited by your card and start being limited by the weekly caps — published as multipliers, so you find the wall by walking into it. A subscription converts a variable cost into a fixed cost and a hard stop. The API has neither. Pick the failure mode you would rather have on a Thursday afternoon.

For what it is worth, the parallel case is the one we build for: a branch, a worktree and an agent session per ticket, which is the part of the problem the terminal multiplexers left unsolved.

Can you use a Claude subscription with third-party agent tools?

Only through Anthropic's own clients. Claude Code's legal documentation states that OAuth authentication is intended for ordinary use of Claude Code and other native Anthropic applications, and that Anthropic does not permit third-party developers to route requests through Free, Pro, or Max credentials on behalf of their users. Building a product? Use an API key.

This is the part of 2026 worth reading rather than reading around. Anthropic's legal and compliance page for Claude Code says OAuth is "designed to support ordinary use of Claude Code and other native Anthropic applications", and that Anthropic "does not permit third-party developers to offer Claude.ai login or to route requests through Free, Pro, or Max plan credentials on behalf of their users."

The distinction that matters is not whether a tool calls itself an agent. It is who is making the call. Running Claude Code on your own machine, under your own login, in a terminal you opened, is ordinary use of Claude Code. Taking that OAuth token and calling Anthropic's servers from a service is not, whatever the service is for.

That is why Fredrin launches the claude binary you installed and logged into rather than asking for a token — and the same for Codex, Cursor and Copilot. The bring-your-own-subscription posture behind our pricing and the Vibe Kanban migration guide survives the rule change because it never held the credential to begin with.

We should concede the other half. The same page notes that advertised Pro and Max limits "assume ordinary, individual usage", and that Anthropic "reserves the right to take measures to enforce these restrictions and may do so without prior notice." Keeping eight agents busy around the clock is not ordinary individual usage. If yours has become a workload rather than a working day, buy API credit for it.

Who should buy Claude Max, and who should stay on Pro?

Buy Max if you hit Pro's ceiling more than once a week, or if you run more than one agent at a time. Stay on Pro if Claude Code is an occasional assistant rather than the way you work. Move to the API if you are building a product on top of Claude rather than coding with it.

• Stay on Pro ($20) if you reach for the agent a few times a week and rarely see a limit warning. The upgrade buys you nothing you are using.
• Max 5x ($100) if you code with an agent most days, or keep a second running while the first thinks. This is the tier the arithmetic above is closest on — which is another way of saying it is hard to get badly wrong.
• Max 20x ($200) if several agents in parallel is your routine and you already know where the 5x wall sits. Buying it first is buying a number you cannot see.
• Use the API if the consumption is programmatic — a service, a pipeline, an SDK integration — rather than you typing. That is the case the terms are explicit about, and where batch and caching discounts start doing real work.

FAQ

Is Claude Max worth it for coding? If you use an agent most working days, yes — it converts a variable bill into a fixed one at $100 or $200 a month. If you code with Claude a few times a week, Pro at $20 covers it and Max buys headroom you will not touch.

What is the difference between Claude Pro and Max? Usage, not features. Anthropic describes Max 5x as five times Pro's per-session usage and Max 20x as twenty times, and both include Claude Code exactly as Pro does. Max also carries two weekly limits, one across all models and one for Sonnet only.

Does Claude Code get its own separate allowance on Max? No. Anthropic's guidance is that Pro and Max usage limits are shared across Claude and Claude Code, so time spent in the web or desktop app draws down the same pool your terminal does.

Can I use my Claude Max subscription with a third-party agent tool? Not by handing it your credentials. Anthropic's Claude Code terms restrict subscription OAuth to its own applications and require third-party products to authenticate with an API key. A tool that runs Claude Code locally under your own login is a different thing from a service that routes requests on your behalf.

Is the API ever cheaper than Claude Max? For light or bursty use, easily — you pay for the turns you take and nothing for the weeks you do not. It stops being cheaper roughly where your usage becomes daily and parallel, and that crossover moves every time Anthropic reprices a model.

Fredrin is a desktop kanban that gives every ticket its own branch, worktree and agent session, running on the Claude, Codex, Cursor or Copilot subscription you already pay for. See what it does.