Cowork
Cowork vs Claude Code
Both let Claude act rather than answer. They differ in where they run, what they touch, and how quickly they consume an allowance — and picking the wrong one is an expensive mistake to repeat daily.
This page is part of the Cowork hub, which covers the whole topic end to end.
The short version: Claude Code is a terminal-native coding agent that works inside a repository. Cowork is a broader working surface for tasks that are not only code — research, documents, multi-step operations across tools. Both are agentic, and both consume far more tokens than chat.
Side by side
| Claude Code | Cowork | |
|---|---|---|
| Where it runs | Your terminal / editor | Claude's own working surface |
| Primary material | A codebase | Documents, data, mixed tasks |
| Best at | Multi-file changes, tests, refactors | Research, drafting, multi-step non-code work |
| Context source | Files it reads from disk | What you provide plus tool output |
| Token profile | Very high — file reads dominate | High — tool output and iteration dominate |
| Supervision | Per-step approval possible | Longer autonomous stretches |
Why both cost so much more than chat
In a chat you send one message and receive one reply. In an agentic run, a single instruction may produce twenty model calls: read a file, plan, edit, run tests, read the failure, edit again. Each of those calls carries the accumulated context of the ones before it. You are paying the quadratic history cost at machine speed instead of human speed.
This is not waste in itself — it is the work being done. It becomes waste when the agent is pointed at a badly scoped task and spends half its calls rediscovering the codebase.
Choosing between them
- Use Claude Code when the answer is a change to files in a repository and you can state the acceptance criterion (tests pass, this behaviour changes).
- Use Cowork when the task spans formats — reading material, producing a document, working across several steps that are not all code.
- Use plain chat when you already know what you want changed and can paste the relevant twenty lines. It is an order of magnitude cheaper and frequently faster.
Keeping agentic spend sane
- Scope narrowly. “Fix the failing test in
auth.test.ts” beats “make the tests pass.” - Give the map. Telling it which files matter saves a discovery phase every run.
- Stop early. An agent that has gone wrong will not right itself; it will re-read everything.
- Restart rather than redirect. A fresh run with a better prompt costs less than steering a long one.
If agentic work is your main use, the upgrade question is genuinely live — see Pro vs Max — but scope discipline is worth more than a tier, and it compounds.