I’ve paid for Claude Code three different ways now: a Pro subscription on my own card, a Max seat expensed through a client, and a two-month stretch of straight API billing during a messy migration. The invoices looked nothing alike. The cheapest line item on paper turned out to be the most expensive option in practice, and the one I resented paying was the one that saved the most hours.
So this is a pricing review written from receipts rather than from a marketing page. Here’s what each tier costs, what it actually buys, and where the maths flips.
Claude Code isn’t sold like a normal coding tool
Most AI coding assistants sell you a seat. GitHub Copilot charges per head, and the model cost is somebody else’s problem. Anthropic does something different. Claude Code is a harness around the Claude models, and what you’re really paying for is token throughput, either bundled into a subscription with usage ceilings or metered per million tokens through the API.
That distinction matters more than the headline price. Two people on identical $20 plans can get wildly different value depending on how they work. A developer who fires off one big agentic refactor at 9am and walks away will hit limits faster than someone doing twenty small, surgical edits across the day. I covered the practical side of that in this working engineer’s verdict on Claude Code, and the pattern holds: usage shape beats usage volume.
The four ways to pay, priced out
Claude Pro: $20 a month
$20 monthly, or roughly $17 a month if you pay annually. It covers Claude Code, the web app, and the desktop client from one pot, but the pot is shared. You get rolling five-hour windows plus a weekly ceiling, and heavy agent sessions will find that ceiling quickly. For a solo developer doing a few focused hours a day, it’s genuinely enough. For anyone running long autonomous tasks, it’s a teaser.
Claude Max: $100 or $200 a month
Max comes in two tiers, 5x and 20x the Pro allowance, priced at $100 and $200. This is the plan most working engineers end up on, and the reason is arithmetic rather than brand loyalty. At API rates, a heavy agentic day on Sonnet 4.5 can run $25 to $60 in tokens once you factor in file reads, tool results, and retries. Four solid days a month and the $200 plan has already paid for itself.
If your usage sits somewhere between casual and hardcore, the $100 tier is the awkward middle child. Plenty of people buy it, blow through the limits in week two, and upgrade anyway. My advice is to skip it unless you know your weekly pattern.
Team and Enterprise: $25 to five figures
Team seats come in Standard (around $25 to $30 per user per month) and Premium (around $150 per user per month), and only the Premium tier includes meaningful Claude Code access. That’s a big jump, and it’s where most procurement conversations get interesting. Our write-up on 30 days of Claude Team with a real workload goes through what that premium actually delivers day to day.
Enterprise is a custom annual commitment, usually negotiated with a floor in the five-figure range. The per-seat cost stops being the meaningful number and total contract value takes over. If you’re weighing that up, the honest cost review of Claude Enterprise breaks the maths down properly, including the parts where the value gets thin.
API: pay per million tokens
Sonnet 4.5 sits around $3 per million input tokens and $15 per million output. Opus runs $15 and $75. Prompt caching takes up to 90% off repeated context reads, which matters enormously for agentic work where the same files get re-read constantly. There’s no minimum, no seat, and no five-hour window. You get a bill at the end of the month that reflects exactly what you did.
Where the subscription quietly wins
Interactive, exploratory work is where flat pricing shines. You’re poking at a bug, the agent reads twelve files, tries three approaches, backtracks twice. On the API you’d feel every one of those detours as a line item. On Max you’d feel them as nothing at all, which changes how you work. People on metered billing self-censor. They write shorter prompts, avoid large-context tasks, and skip the experiments that sometimes produce the breakthrough.
That behavioural tax doesn’t show up on any pricing page, and it’s often bigger than the price difference.
Where the API quietly wins
Automation is the obvious case. CI checks, scheduled refactors, documentation sweeps, batch test generation. These jobs run unattended and don’t need a human in the loop, and a subscription seat dedicated to them is wasted money.
Small teams are the less obvious case. Three developers on Premium seats is $450 a month before anyone writes a line of code. If two of them touch Claude Code twice a week, the API will come in under $100 and nobody has to think about seat allocation. Prompt caching makes this even more lopsided for repetitive tasks.
The costs that never make the pricing page
Every Claude Code bill I’ve seen has a second half that doesn’t appear in the invoice.
- Review time. Generated code still needs reading. Budget 15 to 25 minutes per substantial change, and that’s on a good day.
- Rework. Maybe one in five agentic changes needs a second pass. That’s not a Claude problem specifically, it’s the nature of delegated work.
- Onboarding. Getting a team fluent in prompting, context management, and CLAUDE.md conventions takes a few weeks of reduced velocity.
- Seat sprawl. Premium seats assigned to people who log in twice a month. This is the single most common waste I’ve seen.
- Context upkeep. Keeping project instructions current is unpaid maintenance that quietly determines whether the tool is useful or annoying.
There are also capability gaps worth knowing about before you sign anything long. The limitations review of Claude Enterprise covers the ones that surprise people mid-contract.
A real month, itemised
Here’s my February, which is close to typical. Eighteen working days, roughly four hours of Claude Code use per day, mostly refactoring a Django service and writing tests. About 60% Sonnet, 40% Opus for the harder architectural questions.
On the API, that month came to $214 in tokens, plus $38 for a few CI runs. On Max 20x it would have been $200 flat with limits I never hit. On Pro it would have been impossible after day three. On Team Premium, $150 for a single seat, which sounds cheaper until you remember Premium is designed for organisational features I didn’t need as a solo developer.
The lesson wasn’t that one plan is universally better. It’s that the crossover point sits somewhere around 30 to 40 hours of active use per month. Below that, API or Pro. Above that, Max.
Build your break-even before you buy
Ignore the tier names for a minute and answer three questions: how many hours a week do you actually use it, how much of that work is unattended, and how many people need access. Those three answers map almost mechanically onto a plan.
- Under 10 hours a week, solo: Pro at $20, upgrade the moment you hit limits twice in a month.
- Over 30 hours a week, interactive: Max 20x. It’s the cheapest way to stop thinking about tokens.
- Mostly automation or batch jobs: API with prompt caching, and keep the seat count at zero.
- Five or more developers using it daily: Team Premium, but audit seats monthly.
- Regulated environment needing SSO, audit logs, and a contract: Enterprise, and negotiate the floor hard.
One more thing worth doing before you commit annually. Run a fortnight on the API and actually read the bill. It tells you more about your real usage pattern than any trial period or sales call will, and it costs you nothing to find out. The engineers I know who are happiest with what they pay are the ones who measured first and picked second.