You've seen Claude Code go by — Anthropic's development assistant that works straight from your terminal — and you're asking the only question that matters before you start: how much is this going to cost me every month? The answer fits in one sentence: anywhere from nothing to several hundred dollars, depending on the billing mode you pick and how you use it. There are two modes, the fixed-price monthly subscription and the API billed per token consumed, and they're aimed at completely different people. A beginner coding two hours in the evening and a studio running agents all day long don't pay the same amount, and they don't pay it the same way.
So let's lay out the numbers, the typical bills based on your usage, and the concrete settings that bring the total down without degrading the quality of what you get. By the end you'll know which mode to choose and what budget to expect.
What you're actually paying for when you use Claude Code
You never pay for the Claude Code software itself: it's free to install. What you pay for is access to the AI model doing the work behind it.
This is the part everyone misses at first. Claude Code is a command-line program you install on your computer in about a minute. Installing it costs nothing. But every time it reads a file, thinks, or writes code, it sends a request to Anthropic's servers — and that request is what gets billed.
Billing happens in tokens. A token is a chunk of a word: roughly 4 characters, or about 750 words per 1,000 tokens. You count input tokens (everything Claude reads: your request, your files, the conversation history) and output tokens (everything it writes: the code, the explanations). Output always costs more than input, typically 5 times more.
Worth remembering: a 300-line code file is about 4,000 tokens. If Claude Code reads ten files to understand your project before answering, it has already burned 40,000 input tokens before writing a single line. That's exactly where bills go off the rails — and also exactly where you can cut them in half.
There are three paths to paying for that access. Here they are in the order I'd suggest considering them when you're starting out.
Method 1: learn first, burn tokens second
The cheapest way to use Claude Code is to learn how to drive it before you open the billing tap.
This isn't feel-good advice. A beginner poking around the tool easily consumes three to five times more tokens than a trained user, for a worse result. They resend the same request five times, let the conversation history balloon over hours, let Claude re-read the entire project on every question, and pick the most expensive model to rename a variable. Over a month, the gap between the two adds up to tens of dollars.
The Claude Code in the terminal program from Skilzy is built around exactly that: installation, the commands that matter, structuring a project, keeping consumption under control. You get to the billing question already knowing which plan to take and how not to waste it.
Skilzy is a French AI learning platform covering more than 15 programs (images, video, music, automation with n8n, prompt writing, AI-assisted development). It includes an AI Lab where you use the real tools with credits included, which saves you from stacking three or four subscriptions during your learning phase. Two state-recognized, fundable certifications round it out (RS7439 in AI content marketing, RS6792 in AI and sales).
Be clear on one point: training doesn't replace your access to Claude. You'll still need a Claude subscription or an API key to run Claude Code on your machine. What you save is the month or two of waste that almost everyone pays while learning on the job.
Method 2: the Claude Pro or Max monthly subscription
A subscription is the right call in more than 80% of cases: fixed price, no end-of-month surprises, and Claude Code included.
Anthropic offers several consumer plans, and Claude Code is available from the paid tiers up. Here are the ballpark public prices, in dollars, billed monthly.
| Plan | Indicative price | Who it's for |
|---|---|---|
| Free | $0 | Chatting with Claude on the web, no Claude Code |
| Pro | about $20 per month | Regular but reasonable use, a few hours a week |
| Max (first tier) | about $100 per month | Intensive daily use, several long sessions a day |
| Max (higher tier) | about $200 per month | Very long sessions, several agents in parallel |
These amounts are in dollars and they change. Always check Anthropic's official pricing page before committing — it's the only source that counts.
The mechanics are simple: you pay a flat fee, and you get a generous but capped usage allowance that resets every few hours. If you hit the limit, you wait for the next window or move up a tier. No unexpected charge ever lands on your card.
The right instinct: start with Pro. It's more than enough to learn, build a site, automate scripts, and work on a personal project evenings and weekends. You only move up to Max the day you regularly hit the ceiling — and that day you'll know without any ambiguity, because Claude Code tells you.
For teams, Anthropic offers per-user plans with centralized billing and access management. The price per seat is higher than individual Pro, but the administration is worth it once you're at three or four developers.
Method 3: the API billed per token
The API charges you for exactly what you consume, no more and no less: brutally cheap for light use, dangerously elastic for heavy use.
No flat fee here. You create an API key, top up a balance, and every request draws that balance down. Prices are expressed per million tokens and depend on the model. Here are the ballpark figures for the three tiers:
| Model tier | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Fast, lightweight model | about $1 | about $5 |
| Balanced model | about $3 | about $15 |
| Most powerful model | about $15 | about $75 |
Let's translate that into real usage. A serious work session on a balanced model — say 50,000 tokens read and 10,000 tokens written — costs about $0.15 of input plus $0.15 of output, or roughly $0.30. Thirty sessions like that in a month puts you around $9. At that point the API is clearly cheaper than a Pro subscription.
Now take a developer running the most powerful model on a big project all day, with full code reviews. That same day can cost $20 to $40. Over twenty working days, the bill blows well past Max. That's the whole problem with the API: it protects you from nothing.
Two official mechanisms bring the total down. Prompt caching stores the stable part of your context: writing to the cache costs a bit more, but re-reads are billed at a fraction of the normal price, which pays off handsomely when Claude Code re-reads the same files all day. Batch processing, for non-urgent requests, applies a discount on the order of 50%. The details of these mechanisms and cost tracking are covered in the official Claude Code documentation on costs.
If you go the API route, do one thing before anything else: set a monthly spending cap in your Anthropic console and turn on email alerts. That's two minutes of setup that saves you from a very nasty surprise.
Three sample bills by profile
For the vast majority of people starting out, the realistic budget is around $20 a month, no more.
| Profile | Actual usage | Recommended mode | Monthly budget |
|---|---|---|---|
| Curious, 2 to 3 hrs a week | Small scripts, personal site, learning | API with a cap, or Pro | $5 to $20 |
| Regular developer, 2 hrs a day | One serious project, fixes, tests | Pro | about $20 |
| Full-time developer | All day, large codebase | Max first tier | about $100 |
| Team of 3 to 5 people | Daily production work | Team plans | $60 to $150 per person |
Compare those amounts to what they replace. Twenty dollars a month is the price of one lunch. If Claude Code saves you two hours a week, the math is settled, whatever your hourly rate.
One last useful reference point: pricing is in US dollars. Your bank will apply the day's exchange rate and sometimes a foreign transaction fee. Expect a few percent of difference between the listed price and the amount actually charged.
Six settings that bring the bill down
What costs money isn't Claude Code — it's the useless context you make it re-read on a loop.
Here's what has a measurable effect, in order of impact:
- Clear the history between tasks. The
/clearcommand resets the conversation. Without it, every new message sends all the accumulated history back as input, and you pay several times over for the same information. - Write a
CLAUDE.mdfile at the root of your project. It describes your architecture, your conventions, and your commands once and for all. Claude stops exploring your repo blindly to guess how it's organized. It's the setting with the best effort-to-savings ratio, detailed in our guide on the CLAUDE.md file. - Match the model to the task. Renaming variables or reformatting a file doesn't need the most expensive model. Save that one for architecture and hard debugging, and switch to a fast model for the mechanical work.
- Watch your consumption live. The
/costcommand shows what the current session has cost. Look at it a few days in a row: you'll spot your expensive habits immediately. It's one of the Claude Code commands worth knowing. - Be precise in your requests. "Fix the authentication bug in
auth/login.js, line 42" costs ten times less than "there's a bug somewhere, find it," because Claude doesn't have to comb through the whole project. - Get the plan approved before any writing. Asking for a plan first, correcting it, then launching execution avoids a full rewrite of a feature that went the wrong direction. What you save is the output — the expensive part.
Applied together on the same project, these six points routinely bring a $40-a-month habit down to around $15. Nothing heroic: just less waste.
So, subscription or API?
If you're starting out, take the Pro subscription. Fixed price, zero surprises, and you focus on your code instead of your meter. Move to the API the day you find your usage is either very light and occasional (a few dollars a month) or, at the other extreme, very industrial, with caching and batching to optimize. And if you're still on the fence, install Claude Code, follow a structured training path, and watch your real consumption for two weeks before choosing. The numbers from your own /cost beat any estimate, including the ones in this article.
The prices quoted here are public ballpark figures as of the publication date. Confirm them on Anthropic's official page before committing to anything.
