Last verified

Anthropic · Claude Code

Claude Code.

Anthropic's coding agent · terminal, IDE, desktop and browser · one engine

Claude Code is a coding agent that reads a repository, edits files across it, runs commands and works with git from plain-language instructions. It has no price of its own — access rides on a Claude subscription, a Team seat, an API key or a cloud marketplace. There is no free tier. It is strongest on long multi-file work in an existing codebase, and it is the wrong tool the moment your requirement is offline, fully local, or open source.

Cheapest way in $20/mo
Free tier No
Ways to buy it 8
Oldest check 2026-08-18
§ 01 / WHAT IT IS

What it is.

Short version, before the detail: an agent that edits your repository, not a chat window that suggests code you paste.

The distinction that matters is agency. An autocomplete tool proposes the next few lines and waits; Claude Code takes an instruction, decides which files to open, changes them, runs the commands it needs, and reports what it did. That makes it useful on work that spans a dozen files and useless as a faster way to type — the overhead of describing a one-line change exceeds the change.

It began as a terminal program and now also runs in VS Code and JetBrains, a desktop app and the browser, on one engine that carries your project instructions and settings between them. Which underlying model it uses is deliberately not fixed: it runs the recommended model for your account or your organisation's default, and you override that per session. Configuration is written against aliases rather than model IDs, so it survives the next release.

The parts that show up in practice are the ones about controlling it rather than prompting it — a plan mode that proposes an approach before touching anything, subagents that keep verbose work out of the main conversation, hooks that preprocess data before it reaches the model, and MCP servers for connecting tools it does not ship with. Most complaints about cost and quality trace back to skipping those, not to the model.

Every line below is a question we answered on its own page — most of them by running the thing on our own machine and publishing the transcript. The verdict pill is the answer; the page behind it is the proof. Nothing is repeated here, because a summary of our own evidence is worth less than the evidence.

§ 03 / WHAT IT COSTS

What it costs.

Full pricing →

There is no single price, and any page that gives you one has picked a buyer on your behalf. Four routes, four units of billing — the full breakdown, with the cancellation and refund terms, is on the pricing page.

Route Bought as Price
Claude Pro Subscription $20/mo $17/mo billed annually
Claude Max 5x Subscription $100/mo
Claude Max 20x Subscription $200/mo
Team · standard seat Per seat $25/seat/mo $20/mo billed annually
Team · premium seat Per seat $125/seat/mo $100/mo billed annually
Enterprise Per seat Contact sales
Claude Console API key API key At model rates
Cloud providers Marketplace Set by the cloud
§ 04 / WHEN NOT TO

When it is the wrong tool.

Written from our own runs, including the ones where it refused or fell short. Every item links to the page carrying the transcript, so you can check the claim rather than take it.

You need it to work offline Every turn is an API call. There is no local mode, no degraded offline mode, and no cached fallback — on a plane or behind a firewall that blocks the endpoint, the tool does not run at all. We tested this rather than assuming it. What we ran →
Your code cannot leave the building The CLI runs on your machine, but the reasoning does not. "Local" here means the binary, not the model, and no configuration changes that. If the requirement is that source never reaches a third party, this is the wrong category of tool, not the wrong tier. What we ran →
You need open source Claude Code is proprietary. You cannot read the source, fork it, or audit what it sends. Teams with a policy on this should be looking at the open alternatives from the start rather than negotiating an exception later. What we ran →
You wanted an IDE It plugs into editors; it does not replace one. There is no project explorer, no debugger and no visual diff of its own — it drives the editor you already use. People expecting a Cursor-shaped product are usually comparing the wrong two things. What we ran →
The work is visual It reads images you give it, and it does not make them. Design work, asset generation and anything where the output is a picture rather than a file of code sits outside what this tool does — we ran it and it declined, which is the honest behaviour but still a no. What we ran →
§ 05 / HOW IT COMPARES

How it compares.

All comparisons →

Four pairs we have already worked through in full. Each row says what the pair actually decides, because "which is better" is rarely the question anyone has — the real one is which shape of tool fits the way you already work.

Pair What it decides Verified
Cursor vs Claude Code Terminal agent against an editor. The comparison is about where you want the work to happen, not which model is smarter. 2026-08-13
Codex vs Claude Code The closest like-for-like: two CLI agents from two labs, so this one really is about capability and price. 2026-08-13
Claude Code vs GitHub Copilot Agent against completion. Copilot is priced per seat for a different job; teams often end up paying for both. 2026-08-13
Gemini CLI vs Claude Code The free-tier question. Google's CLI gives away what Anthropic sells, and the trade-off shows up in long sessions. 2026-08-13
§ 06 / ALTERNATIVES

Or something else.

See the list →

If one of the cases above is yours, the list of what to use instead — open source options included — is on its own page, with each competitor's entry price read off the vendor's own site and dated.

Frequently asked.

What people search next.

Q · 01 What is Claude Code in one sentence? +
Anthropic's agentic coding tool: it reads your codebase, edits files across it, runs commands and works with git from plain-language instructions. The full definition covers where it runs and what it is not.
Q · 02 How much does it cost? +
It has no price of its own — access comes with a Claude subscription from $20 a month, a Team seat, an API key billed per token, or a cloud marketplace. Our pricing page lays out all four and says which fits which developer.
Q · 03 Is it worth paying for? +
That depends on whether your work is the shape it is good at — long, multi-file, in a codebase that already exists. For one-line edits the overhead is not worth it. We answered this at length on is Claude Code worth it.
Q · 04 Does it train on my code? +
It depends on your account type and settings, which is why the short answer everyone repeats is wrong in both directions. The detailed answer separates consumer plans from commercial ones.
Q · 05 Which model does it run? +
No fixed one. It uses the recommended model for your account, or your organisation's default, and you override it per session through aliases rather than model IDs — see what model Claude Code uses.
Q · 06 Does it work on Windows? +
Yes, natively — we installed and ran it there rather than reading it off a support page. The test run includes the install command and what the first launch actually does.