.png?w=1024)
AI assistants spent years helping people write emails, summarize documents, and answer questions. Now Anthropic and OpenAI want their AI to do the work itself.
Claude Cowork and ChatGPT Work represent the next phase of that race, giving AI more freedom to work across files, websites, applications, and multistep tasks instead of waiting for users to supply a new prompt at every turn.
The difference matters. Businesses are no longer choosing only which chatbot gives the better answer. They are increasingly deciding how much actual work they are willing to hand over to an AI agent.
- Claude Cowork and ChatGPT Work are chasing the same prize
- The real test is how much supervision they still need
- Browser control changes what an AI assistant can do
- Giving AI more control also creates more risk
- Claude Cowork vs. ChatGPT Work: There isn't one winner yet
- What eWeek found: The better AI may depend on where the work lives
Claude Cowork and ChatGPT Work are chasing the same prize
Anthropic positions Cowork as a way to give Claude a goal and let the AI work across files and tools to complete it. Users can choose the folders and tools Claude can access, watch the steps it takes, and redirect the agent when necessary.
Instead of uploading a single document and asking Claude to summarize it, a user can, for example, grant Cowork access to relevant files and assign a broader task. Claude can work through the material, use available tools, and produce finished work for the user to review.
Anthropic has expanded that idea well beyond local files. Cowork can work across connected tools and the web, and for Pro and Max users, its beta computer-use feature can also interact directly with desktop applications through Claude Desktop on macOS and Windows.
That puts it increasingly close to OpenAI's vision for ChatGPT Work.
OpenAI describes ChatGPT Work as an agent capable of taking action across apps and files, staying with complex projects for hours, breaking goals into smaller steps, and producing finished work. Its browser and desktop capabilities can also help it complete tasks that stretch across websites and desktop apps.
Both companies are effectively trying to remove the prompt-by-prompt bottleneck that has defined generative AI.
The real test is how much supervision they still need
The most useful way to compare Cowork and ChatGPT Work may not be by counting features. It is by asking how much work remains for the human.
Consider a relatively ordinary assignment: research a market, analyze several documents, compare competitors, and turn the findings into a presentation.
A traditional chatbot can help with every part of that assignment, but the user generally remains the project manager. They gather the documents, request the research, check the answers, transfer information between applications, request revisions, and assemble the final deliverable.
Agentic workplace tools are supposed to compress that chain.
Cowork is designed to take a larger assignment and work across the files, tools, and resources users make available to it. That makes it potentially useful for document-heavy workflows as well as for assignments that span browsers, connected services, and other applications.
ChatGPT Work similarly combines research, files, connected tools, web interaction, and artifact creation as it works toward a requested outcome.
The distinction between the two is becoming increasingly difficult to draw from feature lists alone. The more revealing question may be which agent can complete a particular job with the least human intervention.
Browser control changes what an AI assistant can do
The browser is becoming one of the biggest battlegrounds.
On Aug. 26, Anthropic added a built-in browser to Claude Cowork. The browser is rolling out in Claude Desktop on Pro, Max, and Team plans, and on Enterprise plans where an owner has enabled it. When the desktop app is online, the built-in browser can also be available in Cowork on web or mobile.
When a task requires the web, Claude can open its own browser inside Cowork, navigate websites, read pages, click, type, and fill out forms. That browser is separate from the user's personal browser. Anthropic also offers Claude in Chrome for situations where users want Claude to work with a webpage already open in their own browser.
OpenAI has taken a similar approach. ChatGPT Work has a cloud-based browser that can navigate supported websites and perform steps on the user's behalf. The ChatGPT desktop app also has its own built-in browser, while Computer Use can work across permitted apps, tools, and browser tasks on the user's computer.
That means both products increasingly encounter something their chatbot predecessors rarely had to deal with: the messy interfaces humans use every day.
Websites change. Buttons move. Login screens appear. Permissions expire. Pages contain ambiguous instructions. A workflow that looks simple to a person can become surprisingly difficult for an autonomous agent.
That makes reliability more important than an impressive demonstration. An AI that completes nine steps correctly and fails on the tenth may still leave a human responsible for checking all 10.
Giving AI more control also creates more risk
The same capabilities that make Cowork and ChatGPT Work useful create their biggest enterprise question.
An AI that can only suggest an email poses a different level of risk than one that can navigate a website, manipulate files, or take actions on a user's behalf. Anthropic and OpenAI have built permission controls and other safeguards for these agentic systems, but those protections do not eliminate all risks.
Browser-based agents can encounter prompt-injection attacks, in which malicious instructions embedded in webpages or other content attempt to manipulate the AI. They can also encounter unexpected interfaces or situations that require users to review or approve an action.
For businesses, AI adoption becomes as much a permissions problem as a productivity problem.
Organizations will have to decide not merely which employees can use AI agents, but which files, applications, accounts, and actions those agents should be allowed to access.
Claude Cowork vs. ChatGPT Work: There isn't one winner yet
Claude Cowork and ChatGPT Work are increasingly converging on the same destination.
Cowork can work across files and tools, use browser capabilities, create finished work, and continue larger assignments with limited step-by-step prompting. ChatGPT Work can similarly operate across apps, files, websites, and longer projects while producing finished deliverables.
That makes declaring either one a universal winner premature, especially without testing the two agents against the same tasks. Feature lists can tell businesses what an agent is capable of attempting. They reveal much less about how reliably it completes the job. That leads to a different way of measuring the competition.
What eWeek found: The better AI may depend on where the work lives
Claude Cowork and ChatGPT Work are converging on the same goal, but their expanding capabilities make simple feature-by-feature comparisons less revealing. A better question is which agent can take ownership of a particular assignment while requiring the least intervention from the person who handed it off.
If your work looks like this... | Claude Cowork | ChatGPT Work | eWeek's read |
|---|---|---|---|
| “Here are 30 files. Figure them out.” | Designed to work directly across files and folders users make available | Can analyze files and, in the desktop app, work with permitted local files and apps | Both are natural fits; reliability may separate them |
| “Research this and give me a finished deliverable.” | Can research, use tools, and produce finished work for review | Can combine research, apps, files, and artifact creation into finished work | Both are built for this increasingly important use case |
| “Go onto the web and complete this task.” | Built-in browser can navigate sites, read pages, click, type, and fill forms | Browser capabilities can navigate supported websites and carry out multistep tasks | Close contest that may depend heavily on the website and task |
| “Keep working without me prompting every step.” | Designed around receiving a goal and carrying out the task | Can break complex projects into steps and continue working toward an outcome | This is the real battleground |
| “Use sensitive company information to do this.” | Greater access makes permissions, governance, and oversight increasingly important | Broader access to apps, files, and websites creates similar governance questions | Neither gets a free pass on enterprise governance |
| “Do this correctly every time.” | We recommend human review for consequential work | We recommend human review for consequential work | No demonstrated winner yet |
The most important metric may not be how many tasks an AI agent can perform. It may be how often a human has to rescue it. We argue that one of the more revealing ways to measure workplace agents could be the intervention rate: how often a user has to correct the agent, clarify instructions, approve another step, recover from an error, or finish the job themselves.
An agent capable of completing 20 different tasks but frequently requiring human intervention may ultimately save less time than an agent with fewer capabilities that completes its assignments reliably.
That makes intervention rate a useful lens for the next phase of the OpenAI-Anthropic competition. Capability determines what an agent can attempt. Reliability determines how much work businesses can actually hand over.
The company that pushes human intervention as close to zero as possible, without sacrificing accuracy, security, or user control, may ultimately have the stronger workplace agent.
Related reading: As AI moves beyond answering questions, see how enterprises are shifting from AI copilots to autonomous workflows and what that change means for IT leaders.