Claude Cowork vs Claude Code, One Lost Your Laptop

Claude Cowork and Claude Code run the same agentic loop on the same models. Cowork drives documents and connected apps. Claude Code drives a repository, a shell and your git state. Anthropic then changed the terms: since October 6, 2026, new Cowork tasks on Pro and Max run in the cloud.

Which matters more than it sounds. If you picked Cowork to keep work on your own machine, that reason just expired.

Updated October 12, 2026 · by the AITerm team

What separates Cowork from Claude Code?

The audience and the surface, not the intelligence. Both run Claude models through the same agentic loop of gathering context, acting, then checking the result. Cowork wraps that loop in connectors, artifact previews and a folder picker for knowledge work.

In practice, Claude Code wraps it in a terminal with your shell, your git state and your test suite.

So the useful question is never which is smarter. It is which one has hands on the thing you are trying to change.

Look, that table is the whole comparison for most readers. The rest of this article is about the parts that moved in the last four weeks, which no indexed comparison has caught up with yet.

Anthropic is deleting the versus

The framing in every top-ranking comparison is already out of date. Anthropic is merging Cowork into chat and moving its execution to the cloud. Two support articles and one line on the pricing page tell the whole story.

The support article titled Claude Cowork and chat are one Claude says the two are merged into a single conversation, so you no longer pick a mode before starting. The rollout began with Pro and Max on web, desktop and mobile. Once your account moves, you cannot switch back to separate Chat and Cowork options, and Enterprise admins get at least 30 days of notice before anything changes for their organization.

Then the harder change. Honestly, this is the one that should make an engineer pause. The Cowork surface article states that on October 6, 2026, new Cowork tasks on Pro and Max plans run in the cloud, that the "Only on your computer" option in Settings was removed, and that there is nothing to set up and no way back.

Tasks you already started locally stay local until they finish.

Here is the thing. Anthropic's own documentation now points developers one way.

The same support page that announces the cloud migration says that for new tasks which should run only on your computer, use Claude Code in the desktop app, because it runs on your computer and keeps your folders and history there. That is not a subtle hint.

Does Cowork take a lot of tokens?

Yes, more than chat and usually more than the equivalent Claude Code run. Anthropic states it directly in the Cowork guide: each step Claude takes uses tokens, multi-step tasks use more of your usage than a quick question, and auto mode consumes more of your usage limit than the other modes. The Enterprise consumption guide rates Cowork's token intensity above core chat.

That last document has the line engineers should read twice. It notes that agentic work creates intermediate token usage which end users often cannot see.

You are not billed for the answer. You are billed for every read, every retry and every file the agent decided to open on the way.

So we measured it on something real. Our localization audit, checking that every L("key") in the AITerm sources exists in all seven language files, ran once through Cowork in auto mode and once through Claude Code. Same repository, same afternoon, same definition of done.

The ratio sat around three to one against Cowork. We are giving you a rounded ratio rather than a decimal, because this was one task on one account, not a benchmark suite. Honestly, the direction matters more than the digits here.

The reason is structural, not a flaw. Cowork walks a folder tree because it has no index of your project and no cheap way to scope a search.

Claude Code runs a grep. One of those reads nine files, the other reads ninety.

We ported a stalled Cowork task into Claude Code

Look, the October 6 change gave us a reason to test the escape hatch, and we recommend doing it before you need it. Anthropic documents a path out of a Cowork task and into Claude Code. It works, and it took us nine minutes including one avoidable mistake.

# 1. download the zip from the task view, then, outside your project:
mkdir -p ~/cowork-ports/localization-audit
unzip ~/Downloads/cowork-task-export.zip -d ~/cowork-ports/localization-audit

# 2. open that folder as the working directory
claude ~/cowork-ports/localization-audit

# 3. inside the session
/port-task-data

Our mistake is worth repeating because it cost most of those nine minutes. We unzipped into the AITerm repository root, which dumped task artifacts beside Sources/ and left a dirty working tree that looked like a bad merge.

We deserved it. Unzip somewhere neutral, then let the agent read across.

And yes, we noticed the asymmetry. There is a documented command to move a Cowork task into Claude Code.

There is no command to move a Claude Code session into Cowork. Draw your own conclusion about which direction Anthropic expects the traffic to flow.

Is Cowork risky?

Honestly, less than the internet claims, and the risk is not where most comparisons put it. The common framing says Cowork is safe because it runs sandboxed, while Claude Code is dangerous because it has wide system access. That framing was never accurate and it is now backwards in at least one dimension.

Start with what Anthropic documents for Cowork. Claude can only read and write files in folders you have connected. Permanent deletion requires your explicit permission.

In auto mode, Claude reviews each action for safety concerns like data exfiltration and prompt injection, blocks the unsafe ones, and falls back to asking per step if it keeps hitting blocks. Skip mode exists and nothing checks its actions automatically, which the documentation tells you to use only when you fully trust everything involved.

Now Claude Code, which the scary framing tends to describe from memory rather than from the docs. It has four permission modes.

It ships a Bash sandbox on macOS, Linux and WSL2 that isolates filesystem and network access for shell commands, where sandboxed commands get no network by default and any host a command needs is named on the command and reviewed by the classifier. Since v2.1.283, auto mode with that classifier is the starting mode for interactive sessions.

You are probably thinking this is a terminal company arguing for terminals. Fair. So here is the part that cuts against us: for a non-technical colleague processing receipts, Cowork's connected-folder model is genuinely safer than handing them a shell, and we would never suggest otherwise.

So the honest security delta in October 2026 is geography. Cowork tasks on Pro and Max now execute on Anthropic's infrastructure. Claude Code still executes where you start it.

If your threat model has a sentence about code leaving your machines, that difference outranks every permission toggle in this section.

What Cowork still does better than Claude Code

Three things, and they are real. Connectors, previews and scheduling. If your work is documents, data and connected apps, Cowork removes friction that Claude Code would make you build yourself.

Connectors first. Cowork reaches Drive, Gmail and the rest through a connector menu, and respects your network egress permissions, though the documentation notes those permissions do not apply to web fetch, web search or MCPs. Getting the same reach in Claude Code means wiring MCP servers, which is an afternoon you may not want to spend.

Then previews. A deck or a spreadsheet comes back as an artifact you can look at before anything ships. Reviewing a diff is a skill; reviewing a slide is just looking at it.

Scheduling is the one engineers underrate. In practice, Cowork schedules a task with /schedule and the documentation says scheduled tasks run in the cloud with no device online. Matching that with Claude Code means a headless run, a cron entry and somewhere to put the logs.

Turns out the use cases Anthropic lists for Cowork are honest about the target: organizing a Downloads folder, turning receipts into an expense report, synthesizing research, analyzing transcripts, building spreadsheets and decks. Not one of those sentences mentions a branch.

Why a repository still belongs to Claude Code

Because the loop that makes an agent useful on code is build, read the failure, fix, build again. Claude Code closes that loop by itself. Cowork cannot even see it.

Here is the thing. Our build script is the referee. It signs the app and fails loudly on a missing translation key, which means no human has to vote on whether the work is done.

./build.sh
# error: missing localization key "tour.panes.title" in L10n+de.swift
# -> the agent reads this, fixes it, and runs the build again. No prompt needed.

Turns out that one line of output is the entire argument. An agent that can run your build iterates on real feedback. An agent that cannot is asking you to be the feedback, which is exactly the copy-paste treadmill people run into when they try the chat route on a code task.

Branches are the other half. Claude Code knows your git state, so it can work on a branch, show you a diff, use worktrees to keep parallel sessions from fighting over the same files, and run headless in CI. In practice that is where the payoff of multiple agent sessions comes from, and we wrote up what actually paid off in our six worktrees experiment.

Which brings us to the part we care about as a product team. Our anonymous daily ping puts the median user at around three concurrent agent sessions, and the people who get value from that are not watching one agent think. They are reading one while two others work, which is exactly what we built the pane workspace for.

That is a terminal problem more than a model problem. Twenty panes streaming agent output on an M2 Air sit at roughly 140 MB of RAM in AITerm, with keystroke latency under 8 ms, because the renderer is native Swift rather than a browser engine.

Our old Electron setup was past 1.9 GB before lunch. Piloting agents is cheap; watching them in a web stack is not.

How much does Claude Cowork cost?

Nothing on its own. The short version: Cowork and Claude Code both come with a paid Claude plan, and the Free plan includes neither. The choice between tiers is a usage-ceiling choice, not a feature choice.

And yes, the number that should drive your decision is not on that table. It is the shared pool: Anthropic's pricing FAQ states that Claude Code shares the same usage limits as the rest of your plan, so terminal work and chats draw from one allowance. Cowork sits in that same pool and burns it faster than chat.

So the real cost question is not Pro against Max. It is whether you want your weekly allowance spent on an agent walking a folder tree or on an agent running your test suite.

Which one belongs in your week

So, our rule after running both on the same repository: if the work has a branch, a build and a test suite, it goes to Claude Code. Everything else that touches files and connected apps goes to Cowork, now just a mode of Claude. The overlap is smaller than the comparisons suggest.

If you came here deciding what to add, the short version: you already have both, so the decision is only where to spend the pool. Try the port path once while you still have a local task to port. Then tell us what broke: hello@aiterm.app.

Going further

Sources

1. Get started with Claude Cowork, Claude support, 2026: Cowork described as bringing Claude Code's agentic approach to non-coding work, availability on Pro, Max, Team and Enterprise, the statement that Claude can only read and write files in folders you have connected, the manual, auto and skip modes, explicit permission required before permanent deletion, the note that each step uses tokens and that auto mode consumes more of your usage limit, network egress behaviour, plugins bundling skills and connectors, and scheduled tasks running in the cloud with no device online.
2. Use Claude Cowork on web, desktop, and mobile, Claude support, 2026: the statement that on October 6, 2026 new Cowork tasks on Pro and Max plans run in the cloud, the removal of the "Only on your computer" option in Settings, that you cannot switch back, that tasks already started locally stay there until done, that folders stay on your computer and are reachable only through the desktop app while it is open, Team beta and Enterprise admin enablement, and the Download task data then /port-task-data path into Claude Code along with the recommendation to use Claude Code in the desktop app for new local-only tasks.
3. Claude Cowork and chat are one Claude, Claude support, 2026: chat and Cowork merged into a single conversation, the gradual rollout starting with Pro and Max on web, desktop and mobile, existing Cowork tasks, projects, connectors, skills, artifacts and files carrying over, the inability to switch back to separate options once migrated, and at least 30 days of notice for Enterprise admins.
4. Claude pricing, Anthropic, 2026: Free at $0 without Claude Code, Pro at $20 a month or $17 billed annually, Max from $100 with 5x and 20x options, Team Standard at $25 a seat monthly and Premium at $125, Enterprise at $20 a seat billed annually plus usage at API rates, the line stating that Claude Cowork is now just Claude and rolling out to Pro and Max first, and the FAQ statement that Claude Code shares the same usage limits as the rest of your plan so terminal work and chats draw from one pool.
5. Claude Enterprise consumption guide, Claude support, 2026: Cowork's token intensity rated above core chat, and the note that agentic multi-step work creates intermediate token usage that end users often cannot see, so a single Cowork task can use many more tokens than a chat.
6. Choose a permission mode, Claude Code documentation, 2026: the manual, acceptEdits, auto and bypassPermissions modes and what runs without a prompt in each, the classifier as a second model reviewing actions in auto mode, auto mode as the built-in starting permission mode for interactive terminal and VS Code sessions from v2.1.283, the Bash sandbox on macOS, Linux and WSL2 with no network access by default and per-command allowed domains reviewed by the classifier, and the requirement to run --dangerously-skip-permissions only inside a container, VM or sandbox runtime as a non-root user.
7. Claude Cowork is coming to mobile and web, Anthropic, 7 July 2026: the expansion of Cowork beyond the desktop app to web and mobile in beta starting with Max users, tasks continuing when the laptop is closed, scheduled tasks running with no device online, and the statement that nothing ships until you have reviewed and approved it.
8. Manage usage credits for Team and seat-based Enterprise plans, Claude support, 2026: usage credits allowing Team and seat-based Enterprise members to keep working in Claude, Cowork and Claude Code after reaching their included limits.
9. Sandboxing, Claude Code documentation, 2026: filesystem and network isolation for Bash commands, the sandbox modes including auto-allow, and per-command allowed domains in auto mode.