7/16/2026

ChatGPT Work and Claude Cowork: AI apps are becoming a work OS

Realistický cover sjednocené AI pracovní aplikace

For the last two years, AI tools have looked like many separate apps. Chat for questions. Codex or Claude Code for code. Canvas or artifacts for writing. Agents in separate windows. Connectors somewhere in settings. Automations somewhere else.

That direction is changing. On July 9, 2026, OpenAI introduced ChatGPT Work and a new desktop app that brings Chat, Work, and Codex together. Anthropic is building Claude Cowork as a work mode next to chat: an agent that works with files, tools, plugins, scheduled tasks, and longer deliverables. The recent YouTube signals from AI s rozumem and Zaujaloma AI captured the practical issue well: people are no longer asking only “which model is new,” but where the boundary between chat, app, agent, and work environment is going.

That is more important than the redesign itself. AI apps are starting to become a work operating layer.

Diagram: AI apps as a work operating layer

What is new at OpenAI

OpenAI describes ChatGPT Work as an agent for longer, more involved tasks. It can research and analyze information, work across connected apps and files, and create finished documents, spreadsheets, presentations, reports, and Sites. The user can follow progress, answer questions, change direction, and approve important actions.

The new desktop app for macOS and Windows makes the product shift explicit: Chat is for questions and conversation, Work is for research and finished deliverables, and Codex is for software. The former Codex app updates into the new ChatGPT desktop app. The previous ChatGPT desktop app may remain as ChatGPT Classic, but new agent features live in the new app.

One important detail: Work on web and mobile runs in the cloud. Work in the desktop app can use local files and desktop apps with permission. OpenAI also says that at launch, cloud Work conversations do not appear in desktop Work, and desktop Work threads with local files remain on that computer. So the strategic direction is unification, but technically it is not yet one perfectly connected memory.

ChatGPT Sites points in the same direction. A website or lightweight internal app is no longer a separate product on the side; it can be created from Work on the web or from Work/Codex in the desktop app. OpenAI is also retiring Atlas as a separate browser experiment and moving browser-based agentic capabilities into ChatGPT and Codex.

What Claude Cowork shows

Anthropic describes Claude Cowork simply: when you want to hand off a task, switch from Chat to Cowork. Cowork works in selected folders and tools, carries the task end-to-end, and delivers work for review. It runs on web, desktop, and mobile, supports scheduled tasks, and can split larger projects into chunks that run in parallel.

The official product guide makes the practical point clearer: Claude Cowork is for knowledge work, not just programmers. It works with local files, connected apps such as Slack and Google Drive, subagents, long-running tasks, scheduled tasks, and plugins. Anthropic also published data from 1.2 million anonymized Cowork sessions in May 2026. The largest category was business process and operations: reports from scattered updates, onboarding checklists, and spreadsheet reconciliation. The next large category was content creation and copywriting.

That matters. Claude Code is for code. Claude Cowork is for the work around work: reports, decks, checklists, documents, CRM prep, folder audits, and meeting prep. These are exactly the things that consume time inside companies because they require copying context between email, spreadsheets, Slack, CRM, and documents.

Not one super app. Three work modes.

I would not read this as “everything will live in one app.” A better mental model is three modes.

The first mode is Chat. Quick questions, brainstorming, explanation, search, decisions, and short iteration. The output can stay in the conversation.

The second mode is Work / Cowork. Here I do not want only an answer. I want a finished artifact: document, spreadsheet, presentation, report, brief, internal website, onboarding checklist, meeting prep, or CRM summary. The agent needs sources, a plan, intermediate questions, longer runtime, and clear review.

The third mode is agent runtime. Codex, Claude Code, browser, terminal, local files, repositories, tests, and tool calls. Here the agent is not only writing text. It is touching systems. That requires stricter permissions, isolation, audit, and approvals.

If a company does not separate these modes, it will create confusion. People will try to run heavy work in ordinary chat, agents will receive too much access, outputs will not land where they should, and nobody will know what actually happened.

How to deploy this inside a company

I would start with tasks that are expensive because of context switching, not because they require genius. For example:

  • sales meeting prep: CRM, email, call notes, customer website, recent support tickets,
  • weekly status report: Slack, Linear/Jira, GitHub, metrics spreadsheet, comments from team leads,
  • content pipeline: brief, sources, draft, review, localization, publishing,
  • support classification: ticket, customer history, response draft, escalation path,
  • document extraction: contracts, invoices, audit checklist, risks, human approval.

For every process, I would define five things.

First, sources. Which folders, apps, and data types may the agent read? Does it need CRM, Gmail, Drive, Slack, a local folder, or only uploaded files?

Second, actions. Can it only create a draft, or can it write to CRM, send an email, publish a Site, change a file, or create a pull request?

Third, approval. What can happen automatically, and what requires a person? Sending emails, changing customer data, creating legal commitments, and modifying production code should require approval.

Fourth, audit. Store the prompt, sources, output, changes, approvals, and rollback path. Without this, errors and trust will be hard to manage.

Fifth, measurement. Not “how many people use AI”, but whether time to output went down, CRM quality improved, manual copying decreased, support triage became faster, or invoice errors dropped.

Why this is a strong trend

An arXiv study of Codex usage reports that active users of agentic AI grew more than fivefold in the first half of 2026, with the fastest growth outside the original developer audience. More than 10% of users managed three or more concurrent Codex agents in at least one week, and 26.6% used skills to share instructions for more complex workflows.

In other words, agents are not only about programming. Programming was the first good test environment because it has files, tests, git, diffs, and a clear review process. Now the same pattern is moving into knowledge work. Instead of a pull request, the output is a report, deck, spreadsheet, internal website, CRM update, or audit folder.

WorkBench Revisited adds the necessary caution. Workplace agents have improved substantially, but they can still make harmful mistakes, such as emailing the wrong person. So the conclusion is not “let agents into everything.” The right conclusion is: start designing work processes where it is clear what the agent may access, what it must cite, and what a human must approve.

Where this is heading

A year from now, the most important question will not be whether you use ChatGPT, Claude, Gemini, or something else. The more important question will be whether you have a work layer for agents.

That layer will manage:

  • unified access to sources,
  • separation between fast chat and longer work tasks,
  • specialized agents for code, data, documents, support, and sales,
  • company plugins and skills,
  • audit trail,
  • costs and limits,
  • and human approval where an action can break something.

OpenAI and Anthropic are approaching this differently, but the target is similar. Chat is no longer the main product. Chat is the input mode. The value is moving into work agents that can turn conversation, files, and apps into finished output.

For companies, this is good news if they do not treat it as just another desktop icon. It is a chance to redesign a few concrete workflows: less copy-paste, fewer lost notes, fewer manual reports, better audit, and faster follow-up. But only if agents get the right boundaries.

Sources: OpenAI ChatGPT release notes, OpenAI ChatGPT Work and Codex, OpenAI Moving to the new ChatGPT desktop app, OpenAI Creating and managing ChatGPT Sites, Claude Claude Cowork product page, Claude How people are using Claude Cowork, Claude The Claude Cowork product guide, Axios Finding your goldilocks GPT-5.6 model, arXiv The Shift to Agentic AI: Evidence from Codex, arXiv WorkBench Revisited, YouTube signals from AI s rozumem and Zaujaloma AI.