OpenAI Dots vs Claude Cowork: Ongoing Work or Tasks?

OpenAI dots vs Claude Cowork is a decision about how you organise continuing work, not whether one tool can write a document and the other cannot. Both now describe cloud work that can continue while you are away. The distinction is the responsibility you assign, the workspace around it and the controls available in your account.

This comparison uses official public documentation checked on 30 September 2026. It is not a benchmark. For the foundation, what OpenAI dots is explains why an agent is different from selecting a model for a chat reply.

The old desktop-only comparison is out of date

Anthropic’s Cowork getting-started guide now describes cloud sessions in beta, available across desktop, web and mobile. Cloud work can continue with your laptop closed. Local files, a local browser or computer use still require a connected desktop app. That is a different requirement from keeping the device open for every cloud task.

Anthropic’s Cowork and chat merger announcement says the combined experience is rolling out to Pro and Max first, with other plans following. Some accounts therefore still show a Cowork choice while others use a single conversation. Do not judge an account as broken because its interface differs from an older screenshot.

The headline term remains useful for this search: Cowork is the documented work capability, even when the separate mode disappears. What matters in a trial is whether you can start the task, follow it across devices and inspect its result. A menu label is not a measure of intelligence or a guarantee of feature access.

Compare a responsibility with a deliverable

OpenAI’s getting-started documentation for dots presents work as continuing responsibilities that you can correct and reprioritise. A dot can delegate parts to Work or Codex. This can suit a request such as keeping a launch plan current as dates, requirements and feedback change.

The Claude Cowork product page describes handing over goals across chosen files and tools, with progress visible and results ready for review. Think of a task such as turning a folder of exports into a report. This is a useful starting point for evaluation, rather than a restriction that prevents Claude handling recurring work.

There is considerable overlap. Claude’s current documentation includes projects, memory and scheduled work; dots can produce individual deliverables. The honest comparison is therefore about your preferred workflow and evidence from a trial. A statement that one product can “remember” and the other cannot would erase capabilities both vendors now document.

Access and usage are not equivalent packages

OpenAI’s dots access guidance specifies eligible Pro regions, Business Premium access and an administrator-enabled Enterprise beta. A Plus subscription alone is not listed as a dots entitlement. Check eligibility and rollout before comparing paid plans as if both agents were already in the account.

Claude Cowork requires a paid Claude plan according to its getting-started guide. The product page lists Pro at $20 when billed monthly, with annual billing described separately. That is a dated US-dollar plan figure from Anthropic, not a universal local checkout price or a separate Cowork fee.

The same product page warns that Cowork uses limits faster than chat. A longer task can therefore change the value of the plan you already have. OpenAI’s first-dot inclusion also comes with a deeper-work allowance. Do not turn either form of inclusion into a promise that every background assignment will be unlimited.

For a fair cost decision, record which plan you already own and what you would need to add. Then measure your own correction work and the allowance consumed by a representative task. A cheaper headline subscription may still be the right choice, but a price comparison alone cannot establish that it completes your workload.

File access deserves a separate trial

Anthropic’s guide distinguishes cloud work from direct access to files on your computer. Before handing over a folder, create a narrow test folder with copied sample files. Specify whether the originals may be changed. A clear output destination makes it easier to find the result and identify accidental edits.

OpenAI’s computers and apps documentation likewise separates cloud work, optional local-computer access and connected apps. A connected computer must stay online with ChatGPT open for local work. Granting local access is a different decision from asking a dot to prepare a cloud document.

If a workflow depends on a spreadsheet stored on your laptop, test that path explicitly. If all its sources are in an approved shared drive, check the actual connector and account. “Works in the cloud” does not tell you where every source lives, which version is current or who can open the finished file.

A worked example: a monthly operations report

Suppose your job is to prepare a monthly operations report from anonymised exports and a short list of decisions. The output needs reconciled figures, a concise explanation and a draft presentation. The comparison should examine how each system handles changes and uncertainty, not which makes the most attractive first slide.

Give both tools the same copied files and this brief:

Build a draft operations report from these exports. Keep source filenames beside calculated figures. List missing inputs and unresolved definitions. Save a new version rather than overwriting the source files. Draft the presentation only after the figures reconcile. Ask me before sharing anything outside this task.

For a dot, also describe the continuing responsibility: keep the report updated when an approved source changes, and tell you which conclusions need reconsideration. For Claude, place related work in the available project or conversation and explain the same update requirement. You are testing the products you have access to, not forcing either into an artificial limitation.

Introduce a controlled revision after the first draft. Replace a sample export and explain that one category was previously mislabelled. Does the tool rebuild the affected calculations? Does it retain the correction when updating the written explanation? Does the presentation match the revised report? These are concrete checks you can make without publishing confidential business data.

Reject a tidy result if its source figures do not reconcile. Open the calculation behind an important line instead of accepting the narrative around it. Checking an AI answer when you are not the expert provides a method for deciding which assumptions require another person’s review.

Approvals and memory require account-level checks

OpenAI’s controls documentation explains permissions, action review and optional custom rules. Rules are instructions the dot tries to follow, not a guarantee that mistakes disappear. A request to draft a report is not the same assignment as permission to send it to a client.

Anthropic’s merger announcement describes different check-in preferences while keeping the user involved in decisions. Inspect the options in your actual plan before choosing a more autonomous setting. For either product, try an innocuous action that should wait for approval and verify the visible behaviour before expanding the assignment.

Memory can reduce repeated setup, but it can also carry an obsolete assumption into a new report. Keep the authoritative definition in the project brief or source document and correct it explicitly when it changes. OpenAI dots permissions and memory examines the distinction between connection settings and information already retained.

Which workflow should you start with?

Start with Claude’s work capability if your immediate need is a deliverable from selected files and you already have suitable paid access. Start with dots if you have eligible access and want to evaluate responsibility that continues through changing priorities. Both recommendations depend on the workflow, not an unsupported claim that one vendor produces better analysis.

If your organisation’s main requirement is Microsoft 365 ownership and governance, dots vs Copilot Cowork addresses that separate buying decision. For a smaller first assignment, the dots tutorial helps define a result you can actually inspect before expanding access.

Coursium teaches AI skills through lessons and practice on your phone. Its learning approach can help with briefing an agent and reviewing a result; it does not execute cloud tasks or replace these subscriptions. Choose the work you can judge, keep the source files safe, and let a trial of that work decide which tool earns further use.