ChatGPT Work vs Claude Cowork: Inside the Enterprise AI Agent War of 2026

On July 9, 2026, OpenAI began rolling out ChatGPT Work, an autonomous agent that can act across a person's email, calendar, code repositories and messaging apps. It is a significant product, and it is also a late one. By the time it shipped, Anthropic's Claude Cowork had been generally available since April, and Microsoft's Copilot Cowork, built in partnership with Anthropic, had gone worldwide in June. Within a single quarter, the three largest names in applied AI had each put an autonomous "do the work" agent in front of enterprise buyers.

This piece looks at what actually separates these products, why they arrived so close together, and which factors tend to decide adoption once the launch excitement fades. The aim is analysis rather than a verdict: each design reflects a different bet about how AI enters the workplace.

OpenAI

Anthropic

Microsoft

From chatbot to coworker

For three years the dominant enterprise AI interface was the chat box. You asked, the model answered, and a human did the rest. The 2026 wave of "cowork" products is an attempt to move the boundary: the agent is meant to take a stated outcome, break it into steps, use tools, and stay with a task for an extended period without a human in the loop for every action.

OpenAI describes ChatGPT Work as able to take an outcome, decompose it into smaller steps, and stay with complex projects for hours, producing finished documents, spreadsheets, presentations, reports and even websites. Anthropic frames Claude Cowork similarly, as a coworker that reads your files, runs scheduled tasks, and ships work back while you are elsewhere. The vocabulary has converged because the underlying ambition has: less a smarter answer engine, more a junior colleague that operates software on your behalf.

Three products, three shapes

The convergence of ambition hides a real divergence in design. The clearest way to see it is to line the three up by when they shipped, what model runs underneath, and where the agent actually lives.

Claude Cowork is the desktop-native approach. It runs as an application, exposed as a third surface inside Claude Desktop alongside Chat and Code, with file-system access, scheduled recurring tasks, folder-level instructions, a plugin marketplace, and a computer-use capability for driving other software. Anthropic moved it to general availability on April 9, 2026 after public betas on macOS and Windows, and has since layered vertical bundles on top for legal, small-business and marketing-operations work. It is included on paid Claude plans rather than sold as a separate product.

Copilot Cowork is the embedded approach. Microsoft made it generally available worldwide on June 16, 2026, built in partnership with Anthropic, and its distinguishing feature is location: it lives inside Microsoft 365, where a large share of enterprise knowledge work already happens. The pitch is that execution should occur where the documents, mail and meetings already are, rather than in a separate app.

ChatGPT Work is the cloud-agent approach. Rather than anchoring to a desktop or a single productivity suite, it connects outward to Microsoft 365, Google Drive, Slack and Notion, gathers context across them, and runs tasks in the cloud. OpenAI paired it with enterprise governance controls, real-time monitoring and automated red-team security evaluations, and began the rollout with Pro, Enterprise and Edu users before widening to Plus and Business tiers.

The model underneath is a separate choice

One detail in the comparison is easy to miss and important to sit with. Two of the three agents run on Anthropic's models. Copilot Cowork is powered by Claude through Microsoft's partnership, and Claude Cowork obviously runs on Claude, most recently Sonnet 5, the cheaper agentic model Anthropic shipped on June 30, 2026 at an introductory two dollars per million input tokens. ChatGPT Work runs on OpenAI's GPT-5.6.

This matters because it separates two decisions that used to be one. The assistant an organisation standardises on and the model that does the reasoning are no longer the same choice. A company can adopt a Microsoft-branded agent and still be running Anthropic intelligence underneath. That decoupling is the structural story of the enterprise agent market: the interface layer and the model layer are pulling apart, and buyers increasingly touch one without controlling the other.

What actually decides adoption

Launch coverage tends to focus on capability demos. In practice, enterprise adoption of an autonomous agent turns on a narrower set of questions, most of which are not about raw model intelligence.

• Where the work already lives. An agent embedded in the suite a company already runs faces less friction than one that asks users to move. This is Copilot Cowork's structural advantage and the reason ChatGPT Work invests so heavily in outbound connectors to Microsoft 365, Google Drive, Slack and Notion.
• Governance and auditability. An agent that acts autonomously across mail, calendars and repositories is also a new attack surface and a new compliance question. OpenAI's emphasis on monitoring and automated red-teaming, and Anthropic's audit logs and data-retention controls at the enterprise tier, are not incidental features. For a regulated buyer they are the gating criteria.
• The autonomy horizon. "Stays on a task for hours" is a genuine capability claim and also a genuine risk claim. The longer an agent runs unsupervised, the more value it can create and the more a single misstep can cost. How much leash a security team is willing to grant is often the real limiting factor, not the model's benchmark scores.
• Cost predictability. Bundling the agent into an existing subscription, as Anthropic does, reads very differently on a procurement sheet than usage billed per token against an unpredictable workload. The pricing shape can matter as much as the sticker price.

Why they all shipped at once

The near-simultaneous arrival is not coincidence. Three forces converged in 2026. Models became agentic enough to chain many tool calls reliably, which is the precondition for any of this to work. The cost of running those chains fell sharply, with Sonnet 5 explicitly positioned as a cheaper way to run agents and GPT-5.6 marketed on token efficiency, which made hours-long autonomy economically viable rather than a demo. And competitive pressure did the rest: once one credible "cowork" product existed, no major lab could afford to cede the enterprise seat.

The result is that the differentiation has moved up a level. When every vendor can offer an agent that reads your files and executes multi-step tasks, the deciding factors become distribution, trust and integration depth rather than a headline capability that any competitor can match within a quarter.

The lock-in question underneath

For the enterprises doing the buying, the strategic risk is subtler than choosing the wrong agent. It is wiring an entire operating rhythm, the documents, the workflows, the institutional habits, into one vendor's autonomous layer, and discovering later that the model underneath, the pricing, or the availability has shifted in a way that is hard to unwind. The export-control episodes and staggered releases of mid-2026 were a reminder that even a model a company depends on can become temporarily unavailable for reasons outside the vendor's control.

That is why the decoupling visible in the chart above is worth taking seriously as a design principle rather than an accident. The organisations that will weather the churn are the ones that treat the assistant and the model as replaceable inputs rather than permanent commitments, and that keep their work portable across whichever agent and whichever model happens to be best this quarter. Model-agnostic platforms such as Metir AI apply that instinct at the individual and team level, letting people work across the leading models in one place instead of betting everything on a single provider's stack.

The takeaway

ChatGPT Work, Claude Cowork and Copilot Cowork are the same idea expressed through three different theories of distribution: outward-connecting cloud agent, desktop-native coworker, and suite-embedded execution layer. The capability gap between them is real but narrow and shrinking, which pushes the contest toward integration, governance and trust. The most durable lesson for buyers is not which agent won the July news cycle, but that the agent and the intelligence behind it are now separable, and that keeping them separable is the surest hedge against a market still changing shape by the quarter.

Work across every model, not just one agent's

The enterprise agent race shows how quickly the "best" model and the "best" interface can change. Metir AI brings the leading AI models together in one workspace, so your team can pick the right model for each task and keep its work portable as the landscape shifts. Try Metir AI free and stay flexible while the agent war plays out.

Sources:

• OpenAI introduces ChatGPT Work, a cloud-based AI agent that manages tasks across email, Slack and calendars | VentureBeat
• OpenAI launches ChatGPT Work, deepening race for workplace AI tools | BNN Bloomberg
• OpenAI Debuts ChatGPT Work Workplace AI Agent With GPT-5.6 | Forbes
• Anthropic launches Claude Sonnet 5 as a cheaper way to run agents | TechCrunch
• Introducing Claude Sonnet 5 | Anthropic
• OpenAI's ChatGPT Work Promises Enterprise Automation, But Will Leaders Trust It? | UC Today