What Claude Code is — and how it differs from asking a chatbot
Claude Code is a command-line coding agent built by Anthropic. Rather than having a developer paste code into a chat window and copy an answer back, it reads the project folder directly, edits files, runs commands and checks the results
The three lines
- Form — a tool that runs in the terminal. Given a natural-language instruction, it reads, edits and executes
- Difference — a chatbot tells you what to write; an agent does the writing. One instruction triggers dozens of model calls
- Scale — Anthropic said revenue from it grew 5.5-fold through July, and it is cited behind the Q2 2026 revenue reversal
Key questions
- What is Claude Code?
- A coding agent built by Anthropic. It first appeared in February 2025 and was made generally available that May. What distinguishes it is where it runs: in a terminal rather than a browser chat window. A developer launches it inside a project folder and says something like 'add password reset to the login screen.' The tool then finds and reads the relevant files, decides what to change, writes the code, and runs tests to check the result. A web version and sandboxing arrived in October 2025.
- How is that different from asking ChatGPT about code?
- In how much of the work the person does. Asking a chatbot means finding the relevant code, copying it, writing a question, pasting the answer back into an editor, running it, and pasting any error back in again. A coding agent does that loop itself: find, read, edit, execute, check, repeat. That is why a single instruction triggers dozens of model calls. The person moves into the reviewing seat instead. Since the tool can make wrong edits, review remains a human job.
- Why should a non-developer care?
- Because it changed the revenue structure of the AI industry. In the second quarter of 2026, Anthropic's revenue reached 11.6 billion dollars against OpenAI's 6.7 billion — the first such reversal — and reporting consistently cited this product. For three years AI companies were valued on how many consumers they gathered. Coding agents are bought by companies at team level and call the model dozens of times per task. In a market where model prices keep falling, doubling revenue requires usage to grow faster than price drops, and this product's structure produces exactly that.
- Does this mean developers are no longer needed?
- Not on what is currently confirmed. The tool executes instructions; deciding what to build and judging whether the result is correct remain human work. It sometimes makes wrong edits, so review is required — and reviewing means being able to read code. Enterprise case studies emphasise throughput per person rather than headcount reduction. That said, this describes the position as of August 2026, and as the tools improve the human role moves with them.
In the second quarter of 2026, Anthropic's revenue passed OpenAI's for the first time: 11.6 billion dollars against 6.7 billion.
What reporting pointed to was not model quality but product shape. The product is Claude Code.
1. What it is
| Item | Detail |
|---|---|
| Maker | Anthropic |
| First release | February 2025 (general availability May 2025) |
| Where it runs | The terminal; a web version added October 2025 |
| What it does | Reads, edits, executes and checks project files |
| How you instruct it | Plain language |
The second-to-last row is the point. It does not stop at reading. It edits and runs.
That is what "agent" means here: a program that takes multiple steps on its own until the instruction is complete.
2. The difference from a chatbot — who does the loop
The same task, two ways:
| Step | Asking a chatbot | Instructing an agent |
|---|---|---|
| Find the relevant files | Person | Tool |
| Copy and paste code | Person | None — it reads directly |
| Write the question | Person | Person (one sentence) |
| Move the answer to the editor | Person | None — it writes directly |
| Run it | Person | Tool |
| Paste the error back | Person | Tool (loops on its own) |
| Review the result | Person | Person |
The human column shrinks from six rows to two. The four that left did not disappear; they moved.
What did not move is the last row. The tool can make wrong edits. Deciding what to build, and judging whether the output is right, stays human.
3. How that shape produced revenue
This is where the story reaches people who do not write code.
Every cell marked "Tool" in the right-hand column is a place where the model gets called. Once to read the files, once to decide what to change, once to write, once more to read an error and try again.
One human sentence produces dozens of model calls.
| Consumer chatbot | Coding agent | |
|---|---|---|
| Model calls per instruction | 1–2 | Dozens |
| Who signs the contract | An individual | A company (team, then department) |
| Deal value | Tens of dollars a month | Hundreds of thousands to millions a year |
| Cancellation | One click | Runs to contract end |
| How it spreads | Person to person | Team to department to company |
As this page set out on August 10 in "What AI token pricing is," model prices have fallen steadily for half a year. Gemini 3.7 Flash halved its price on August 13; DeepSeek raised prices and was treated as the exception that proved the direction.
In a market where the unit price is falling, doubling revenue requires usage to grow faster than price drops. A product that calls the model dozens of times per instruction builds that condition into its structure.
Anthropic said revenue from the product grew 5.5-fold through July. How much of the 11.6 billion dollars it represents has not been disclosed.
4. What it does well and badly
The following summarises public materials and user accounts. This page ran no testing of its own.
| Type of work | Relative strength |
|---|---|
| Mechanical edits across many files | Strong (renames, formatting) |
| Reading and explaining existing code | Strong |
| Writing tests and tracing errors | Strong |
| Deciding what to build | Human work |
| Judgements whose context is outside the code | Weak |
| Deciding whether the result is right | Human work |
A pattern emerges. It is good at work whose answer is inside the project and bad at work whose answer is outside it.
Questions answerable by reading every file — where is this function used, what breaks if I rename this variable — the tool handles faster than a person. "Would a user actually want this feature" has no answer anywhere in the codebase.
5. Common questions
Q. Can a non-developer use it? Installing and running it is not hard. Reviewing the output requires reading code. Without review, a wrong edit goes unnoticed.
Q. Does my code leave the company? It uses a cloud model, so portions of code are transmitted. Enterprise agreements set separate data-handling terms. Check internal policy first.
Q. How does it compare with Cursor or GitHub Copilot? They are products in the same category, differing in form: inside the editor (Copilot), as the editor (Cursor), in the terminal (Claude Code). Benchmark comparisons vary enough that this page states no ranking — the problem covered on August 12 in "What a coding benchmark is."
Q. How is it priced? Both subscription and metered API access exist. Under metering, cost scales with call volume, so complex work on a large project adds up quickly.
6. What could not be confirmed
- Revenue contribution — no disclosure breaks out the product's share of 11.6 billion dollars. The attribution is a common reading, not a verified decomposition.
- The 5.5-fold figure — Anthropic's own, with no disclosed baseline or methodology.
- Performance — untested here. The strengths table summarises public materials and maker documentation.
- Competitor comparison — benchmark results diverge, so no ranking is given.
- Related — the revenue reversal is covered in "Anthropic posts $11.6bn in Q2 revenue," and pricing structure on August 10 in "What AI token pricing is."