Claude Code ships software from a terminal, while Claude Cowork builds spreadsheets, decks and reports from your files. This breakdown sets the two side by side on interface, safety, pricing across every plan, shared usage limits and the jobs each one handles best.
Two Anthropic tools, one subscription, and a lot of confusion about which one to open. Claude Code is the agent developers run in a terminal to ship software. Claude Cowork is the same agent idea pointed at folders, spreadsheets, slide decks and research, built for people who never touch a command line. In mid-September 2026 the picture got blurrier on purpose: Anthropic folded the Cowork tab into the main Claude chat, so Claude now decides when a request needs a longer agentic job.
This article sorts out the overlap. You will see how the two tools differ in daily use, what each one costs on every plan, and which jobs belong to which tool. Prices and rollout details come from Anthropic's own pages and recent reporting, and anything still moving is flagged as such.
💡 Short answer: if your output is a pull request, use Claude Code. If your output is a spreadsheet, a memo or a slide deck, use the Cowork side of Claude. If you do both, one paid plan includes both and draws from a single shared usage pool.
Claude Code lives where developers already work. You can run it in a terminal, inside VS Code or JetBrains IDEs, from the Code tab of the desktop app, on the web at claude.ai/code, and on mobile. You describe a goal in plain English. It reads the repository, plans the change, edits files across the project, runs commands and tests, then reports back with a summary of what it did.
A typical session sounds like this: "Add rate limiting to the signup endpoint and write tests for it." Claude Code finds the route files, proposes a plan, edits the code, runs your test suite, fixes what fails and leaves you a diff to review. Its outputs are the artifacts engineers care about: commits, pull requests, tests and refactors. By default it works on your own machine, so it sees the same files, tools and git history you do. A CLAUDE.md file in the repository gives it standing instructions, and the tool supports skills, hooks, subagents, MCP servers, plugins and an Agent SDK for anyone who wants to build custom workflows on top of it.
Claude Cowork launched in January 2026 as a research preview aimed at people who want the agent behavior of Claude Code without writing code. Instead of typing commands, you hand it a task and a folder: "Reconcile October receipts against the bank export and flag anything that does not match." It opens the files, plans the steps, works through them and hands back a finished workbook with a mismatch tab and a short summary.
The outputs look nothing like a diff. Expect Excel files with working formulas, PowerPoint decks, formatted documents and memos. According to Anthropic's support documentation, Claude's work runs in an isolated environment on Anthropic's servers, and when a task needs a local file or your browser, Claude reaches it through the Claude Desktop app. Cowork is available on the desktop app, on the web, on mobile and in a Chrome side panel, depending on your plan, and it asks for explicit permission before permanently deleting any file.
On September 16, 2026, Anthropic announced that Cowork and chat are merging into one experience. The Cowork tab is going away, and Claude now works out what each request needs instead of asking you to pick a mode. In Anthropic's words: "We stopped making you choose." The rollout reaches Pro and Max subscribers first across web, desktop and mobile over several weeks, with Team and Free plans following and Enterprise administrators getting at least 30 days' notice. Claude Code stays a separate tool, and Docs and Slides launched in beta alongside the change.
In practice, "Cowork" is becoming less of a product you open and more of a set of abilities inside Claude. The comparison with Claude Code still holds, because it was never really about two buttons. It is a comparison between two kinds of work: building software and producing business documents.
💡 If your app no longer shows a Cowork tab, you are not missing a feature. The abilities moved into the main conversation, and your existing skills, connectors and projects carry over.
The fastest way to see the split is a table. Both tools run the same kind of agent loop: plan, act, check and repeat. What changes is the surface, the safety model and the type of output.
| Area | Claude Code | Claude Cowork |
|---|---|---|
| Built for | Software engineers | Operations, research, finance and legal teams |
| Main surface | Terminal, IDEs, desktop Code tab, web, mobile | Desktop app, web, mobile, Chrome side panel |
| Works on | Repositories, git history, CI pipelines | Documents, spreadsheets, local folders, the browser |
| Typical output | Commits, pull requests, tests, refactors | Spreadsheets with formulas, decks, memos |
| Standing instructions | CLAUDE.md and auto-memory | Global and folder instructions, project memory |
| Extensions | Skills, hooks, subagents, MCP servers, plugins, Agent SDK | Plugins, connectors, sub-agents |
| Automation | Routines, desktop scheduled tasks, /loop | /schedule for recurring cloud tasks |
| Approvals | Tool permissions, Plan Mode, hooks | Manual, Auto or Skip modes |
Claude Code assumes comfort with a terminal. By default it asks before it edits files or runs commands, and you can switch to Plan Mode for a read-only review that proposes changes without making them. Developers who want tighter control add hooks that run their own checks before or after each action.
Cowork hides that machinery. You install connectors from a browse-and-click marketplace instead of editing a config file, and you pick an approval style: Manual, Auto or Skip. The more freedom you grant, the more you rely on the agent reading content safely. Web pages and documents can contain text written to hijack an agent, a risk known as prompt injection, so keep approvals on for anything touching files you cannot replace.
Both tools remember how you like to work. Claude Code reads CLAUDE.md files and also reads AGENTS.md, which is handy for repositories shared with other coding agents. Cowork keeps global and folder instructions plus project files with memory, and that memory is now shared with chat.
Connectors differ in flavor. Claude Code leans on MCP servers, git hosting and CI systems such as GitHub and GitLab. Cowork leans on the business apps where documents already live, bundled as plugins that combine skills, connectors and sub-agents for a role like finance or legal.
Neither tool is sold on its own. Both come with a paid Claude plan, so the real question is which plan fits how much agent work you run.
| Plan | Price | Claude Code | Cowork abilities |
|---|---|---|---|
| Free | $0 | No | No |
| Pro | $20 per month, or $17 per month billed annually ($200 upfront) | Yes | Yes |
| Max 5x | $100 per month | Yes | Yes |
| Max 20x | $200 per month | Yes | Yes |
| Team | About $20 to $25 per seat per month | Yes | Yes |
| Enterprise | Quoted by Anthropic sales | Yes | Yes |
Max plans add a 5x or 20x usage multiplier, higher output limits and early access to new features, and they are billed monthly only. Anthropic changes plans often, and the Cowork rollout to Team and Free plans is still in progress, so check Anthropic's pricing page before you commit.
A little quick math helps:
💡 Budget tip: start on Pro for a month and watch Settings > Usage. If you hit the five-hour limit every day, Max 5x is the logical next step.
Here is the part most comparisons skip: Claude Code and Cowork draw from one shared usage pool. Anthropic describes it this way: every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top. A long Cowork job and a long Claude Code session spend the same budget.
Agent work spends that budget faster than chat. Anthropic's documentation notes that multi-step tasks use more of your usage than a quick question, because each step Claude takes uses tokens. When you reach a limit you can wait for the reset, move to a higher plan, or turn on usage credits to keep working at standard API rates. Grouping related work into a single task, instead of many small ones, also helps.
Developers have one more option. Claude Code can run against an API account billed per token, which suits CI pipelines and teams that want metered spending instead of plan limits.
Claude Code earns its place when the work lives in a repository and success means code that runs.
/loop to run dependency updates, changelog drafts or triage on a schedule.
Claude Code is also the better pick when you need control. Plan Mode, hooks and per-tool permissions let a team set rules about what the agent may touch, which matters on a production codebase.
Cowork earns its place when the work lives in files and websites and success means a document someone can open.
💡 Write your folder instructions once, then reuse them with
/schedulefor recurring cloud tasks like a weekly report.
If your week is repositories and pull requests, Claude Code is the default and Cowork is an occasional helper for reports. Many developers use both on one Pro or Max plan: Claude Code for the build, Cowork for the release notes, the stakeholder spreadsheet and the investor update.
| If your task is... | Pick |
|---|---|
| Fixing a failing test suite | Claude Code |
| Building a quarterly revenue workbook | Cowork |
| Migrating code between languages | Claude Code |
| Turning 40 PDFs into a briefing | Cowork |
| Reviewing a pull request | Claude Code |
| Sorting a messy shared folder | Cowork |
| Drafting a deck from a data export | Cowork |
| Scheduling a nightly dependency check | Claude Code |
If you work in operations, finance, research or legal, start with Cowork. You never touch a terminal, you work in files you already have, and the output arrives in formats your team opens daily. Keep approvals on Manual while you build trust, and review every formula and every cited clause yourself. An agent can assemble a contract summary in minutes, but the signature still belongs to you.
Freelancers and small shops usually need both. A consultant might use Cowork to build client reports, then Claude Code to automate the repetitive parts as a script. A single Pro plan includes both, which keeps the stack simple while you find your usage level.
A team of five can move to Team seats, which add project sharing, while Enterprise administrators get at least 30 days' notice before changes like the Cowork merge reach their workspace.
Claude Code and Cowork are Anthropic's own apps, and they are not available on PicassoIA. The Claude models are, and a plain chat window handles plenty of the same drafting and coding work: a bug explanation, a spec, a first-pass script, a summary of a long document. Claude Sonnet 5 is the best place to start, with Claude Opus 4.7 and Claude 4.5 Haiku as alternatives for deeper reasoning or faster replies.
Open the Claude Sonnet 5 page, sign in and type your request in the prompt box. You can paste a stack trace, describe a feature in plain English or attach a screenshot of an error, since the model reads images.
| Parameter | What it does | Default |
|---|---|---|
| Prompt | Your request (required) | None |
| Effort | Thinking level: low, medium, high, xhigh or max | low |
| Max tokens | Caps the length of the response | 8192 |
| System prompt | Fixes a role or coding style for the session | Empty |
| Image | Optional screenshot or diagram | None |
| Max image resolution | Scales images down to save time and cost | 0.5 megapixels |
Use low for fast one-line fixes and high or max for a bug that touches several files. A short system prompt such as "You are a senior reviewer who explains every fix in plain language" keeps the tone steady across a project.
Paste the output into your project, run it, and send back any error message. Keep the same system prompt for a whole project. For document work, ask for an outline or a table first, then expand each section one at a time.
Every photograph in this article was generated with p-image on Picasso IA, from prompts that name the lens, the light and the texture of the scene. That is the real trick: specific prompts give photographic results. Try the same idea with GPT Image 2, Seedream 4.5 or Flux 2 Pro and compare how each one reads the same prompt.
Start with something from your own work: a product on a desk, a team meeting, a blog header. Write one sentence for the subject, one for the light and one for the camera, then generate. Change a single detail and generate again.
Open Picasso IA, pick a text-to-image model and make your first image today. Then keep experimenting with a new angle, a new light and a new lens.
