Claude Code requires a paid Claude subscription or an Anthropic API account. Pro is $20 a month, Max is $100, Max 20x is $200, and direct API access bills per token at Sonnet and Opus rates. There is no free tier. At Formaum I run Max for daily client work and fall back to the API for bulk agent workloads. The right tier is not the cheapest one. It is the one that matches how you actually work.
Most pricing posts compare the tier numbers and stop there. That is not useful. The question is not what each plan costs. The question is what each plan delivers per hour of real engineering work, and at what point you stop paying for software and start paying for friction.
The actual Claude Code subscription landscape
There are three ways to pay for Claude Code. Subscription, API, or both.
Pro at $20 a month. Includes Claude Code in the terminal, in your IDE, and on the web. Both Sonnet and Opus are available. Usage runs on a rolling five-hour window, roughly 44,000 tokens and 45 messages per window. Hit the cap at 2pm and you wait until 7pm.
Max 5x at $100 a month. Five times the Pro allowance in the same five-hour window. Roughly 220,000 tokens. Built for developers running multi-hour sessions with full codebase context loaded.
Max 20x at $200 a month. Twenty times the Pro allowance. Roughly 440,000 tokens per window. The tier where rate limits stop being a thing you think about.
API access. Pay-per-token through console.anthropic.com. Sonnet and Opus billed at published per-million-token rates. No monthly commitment, no rolling window, no caps you cannot lift by adding credits.
Team plans exist at $20 per seat per month for Standard and $100 per seat for Premium, with Claude Code only included on Premium. Enterprise bills seat plus API usage. If you are a solo or two-person shop, ignore those. The decision is Pro, Max, or API.
What you get at each tier
The feature surface is the same across tiers. Sonnet, Opus, terminal, IDE, web, MCP servers, hooks, custom slash commands, plan mode. What changes is how long you can run before the door closes.
Pro gives you maybe ninety focused minutes of agent work per five-hour window before you start hitting limits on a real codebase. Max 5x gets you through a full half-day. Max 20x lets you run multiple parallel sessions or full-day refactors without watching the meter.
API gives you no cap at all. You pay for what you use. The tradeoff is that you are now writing checks per task instead of per month, and that math gets ugly fast on big context windows.
How I think about Claude Code cost
The wrong question is what does it cost per month. The right question is what does it cost per shipped task.
If Claude Code saves me two hours on a build, the marginal cost of that build session is somewhere between zero and a few dollars of token spend. Compared to my client hourly rate, the subscription is rounding error. The math is not about the $200. The math is about the hours.
I run two filters before picking a tier. First, am I hitting the cap on Pro? If yes, upgrade. Second, am I using less than half of Max? If yes, downgrade. There is no loyalty discount for over-subscribing. Anthropic does not care that you paid for unused tokens last month.
For client-facing engineering work I default to Max 5x. For pure experimentation and weekend builds I stay on Pro. For agent workloads that run unattended and burn through context, I use raw API access and route the cheap reasoning tasks to Haiku instead of paying Opus rates for grep.
When Pro is enough
Pro covers most working developers comfortably. Use Pro if you are doing one or two focused Claude Code sessions a day, your codebase is under fifty files, you mostly work in single repos, and you are willing to wait out the occasional rate limit. That is a real working setup. It is not a limitation.
If you are still learning Claude Code, stay on Pro for at least a month before upgrading. Most of the people I see complaining about Pro limits are burning context on bad prompts, not real work. Fix the prompts first.
When Max pays back
Move to Max 5x the first month you hit Pro rate limits more than twice. The math is straightforward. If hitting the cap costs you an hour of waiting, and you bill or build at any meaningful rate, $80 of incremental subscription cost buys back hours of throughput. That is a trivial trade.
Move to Max 20x if you run multiple Claude Code sessions in parallel, work on large monorepos where context regularly exceeds 100,000 tokens, or use Claude Code as the build layer for client work where downtime translates directly to slipped deadlines. At $200 a month, Max 20x is cheaper than one billable hour at any professional engineering rate.
When raw API access wins
API beats subscription in three cases. First, batch agent runs. If you have a Trigger.dev pipeline or a cron job spawning Claude calls overnight, that is not subscription work. Pay per token, route cheap tasks to Haiku, and your unit economics actually make sense.
Second, intermittent use. If you only touch Claude Code three days a month, the subscription is wasted. Buy $20 of API credit and pay for what you use.
Third, custom integrations. Anything you are wiring into your own infrastructure, your own CRM, your own backend, lives on the API. Subscription tiers are for the interactive client. The API is for everything you build yourself.
In practice, most serious operators end up running both. Subscription for the interactive sessions, API for the automated workloads. They are not competing. They are different surfaces of the same platform.
The token-burn habit, and how to stop it
Most of the people complaining about Claude Code cost are not paying for Claude Code. They are paying for sloppy prompting.
Three habits cut token spend by half without changing tiers. Clear the session between unrelated tasks. A thirty-message context full of yesterday's debugging is dragging every new prompt down. Type /clear and start fresh.
Reference specific files instead of letting Claude search. Typing @src/lib/webhook.ts is cheaper than asking Claude to find the webhook handler. Cheaper and faster.
Default to Sonnet for execution. Use Opus only when you need genuine reasoning. Opus is roughly five times the cost of Sonnet per token. Most coding tasks do not need Opus. Most strategy and architecture tasks do.
Do those three things and the rate limit conversation stops being relevant. You can run on Pro for work that would otherwise blow through Max.
The Claude Code subscription is not the bottleneck. The way you use it is. Pick the tier that matches your actual usage, default Sonnet over Opus, clear sessions between tasks, and the cost question disappears. The infrastructure that makes an AI system actually run on a Tuesday at 3am when nobody is watching does not get built on whichever tier was cheapest. It gets built by people who match the tool to the work.
Run on a stack that's holding you back?
Book a 45-minute discovery call. I'll map what moves, what stays, and what makes sense for your operation.
Book a callFrequently Asked Questions
Genevieve Claire
Builds bespoke AI automation systems for multi-location operations. Previously EA Sports FIFA ($7B franchise) and Film/TV VFX on Skyfall, Avengers, Game of Thrones. Based in Vancouver, BC.