On July 9, 2026, OpenAI overhauled its Windows desktop app, merging its familiar ChatGPT interface with the developer-centric Codex coding environment and a new autonomous agent called Work. The update also introduces the GPT-5.6 model family—Sol, Terra, and Luna—giving users more control over speed, reasoning depth, and cost than ever before. But for Windows users, the crowded feature set raises an immediate practical question: which mode and model should you actually use for a given task?

One App, Three Modes: Chat, Work, and Codex

The new ChatGPT desktop application, available globally on Windows and macOS, is no longer just a chat window. It now presents three distinct modes:

  • Chat: The classic conversational interface for everyday prompts, questions, and brainstorming. It remains the simplest way to interact with the underlying language models.
  • Work: A long-running, multi-step agent capable of research, analysis, document creation, spreadsheet generation, presentations, and even building simple websites. It can access local files and desktop applications with explicit user permission and includes a built-in browser that supports tabs, downloads, and authenticated sessions.
  • Codex: A full-fledged coding environment with project management, repository integration, terminal access, and developer tools. Existing Codex users can update to the new app without losing projects or settings.

OpenAI has rebranded the previous standalone ChatGPT client as "ChatGPT Classic," which will continue to receive security updates and bug fixes but will not gain these new agent features. For Windows users, the desktop version of Work is particularly significant because of its ability to interact with local resources—a capability that goes well beyond what the web interface offers. However, that power comes with the responsibility to manage permissions carefully, treating the agent like any automation tool that can read, write, and browse on your behalf.

A New Model Lineup: Sol, Terra, Luna, and When to Use Each

Alongside the mode merger, OpenAI introduced the GPT-5.6 family, comprising three models tailored for different needs:

  • Sol: The flagship model, built for maximum capability and reasoning. It is available in Work and Codex for paid subscribers (Plus, Pro, Business, Enterprise) and can be accessed in standard Chat only through higher reasoning settings on eligible plans.
  • Terra: The balanced, everyday option. It occupies a middle ground between power and speed and is the default for Work and Codex for free and Go users, while paid users can select it manually.
  • Luna: The fastest and most cost-effective model, ideal for simple, high-volume tasks. Paid users can choose Luna in Work and Codex to conserve usage allowances.

In ordinary Chat conversations, the model selection is more limited: GPT-5.5 Instant remains the default for quick responses, and Terra and Luna are not selectable there. Reasoning effort settings now let users dial up how long the model thinks before responding—higher effort can improve complex analysis or coding but consumes usage quotas faster and increases latency. OpenAI’s own advice is pragmatic: start with the lowest effort setting that might work, and escalate only if the output isn’t good enough.

What It Means for Windows Users, Developers, and IT Admins

For Everyday Windows Users

If you use ChatGPT mainly for writing, research, or casual help, you can stick with Chat mode and not worry about the other features. But the Work agent is where the real productivity gains lie—if you’re comfortable with the permissions. Imagine asking it to gather data from a spreadsheet, generate a report in Word, and email it to a colleague, all while you step away. The built-in browser can even log into sites on your behalf. The catch: you must explicitly grant these permissions, and you should only do so for tasks you trust. Treat Work like a macro recorder that can make mistakes; review its actions, especially early on.

For Developers

Codex users gain access to a unified environment that now hosts all their projects alongside the new models. Sol’s enhanced reasoning can untangle intricate codebases, while Luna’s speed is perfect for quick edits or boilerplate generation. The reasoning slider is a welcome addition: for a tricky debugging session, crank it up; for autocompleting a simple function, dial it down. Note that Codex is only available on desktop—mobile and web users cannot activate this mode, though some desktop tasks can be triggered remotely from the ChatGPT mobile app’s Remote tab.

For IT Administrators

The most critical step is checking whether Work is enabled for managed workspaces. OpenAI has not enabled it universally by default, so you can decide whether your organization’s data should flow through an agent that can read local files and interact with desktop apps. Because desktop Work threads and local file access remain confined to a single computer—cloud-based Work conversations do not yet sync to the desktop—you might need to set policies on whether sensitive data can be processed locally. Communicate clearly with staff about the approved use of Work and the risks of granting broad file permissions.

How We Got Here: The Rapid Evolution of ChatGPT on Desktop

OpenAI’s Windows journey started with a basic chat app, followed by the integration of Codex as a separate environment. This latest release collapses those silos and adds an agentic layer that reflects the industry-wide pivot toward autonomous AI assistants. Microsoft’s own Copilot has been pushing deeper into Windows, and tools like Google’s Bard have blurred the lines between chat and agent. OpenAI is responding not just with more models—a trend that began with GPT-4 and its variants—but with a mode-driven interface that acknowledges users need different tools for different jobs. The introduction of reasoning controls is a direct answer to the criticism that advanced models are overkill for simple tasks, yet underpowered when you need depth.

What to Do Now: Practical Steps for Every User

  1. Update strategically: If you want the new features, install the new ChatGPT app. If Chat is all you need, the Classic version remains stable and supported.
  2. Audit Work permissions: Before you let Work access local files or apps, start with sandboxed tasks. Use the built-in browser only for sites you’re willing to let an AI interact with.
  3. Pick the right model: For a quick draft email, Luna is plenty. For analyzing a messy dataset, choose Sol with moderate reasoning. Don’t default to the strongest model out of habit—you’ll burn through your usage cap.
  4. Leverage reasoning settings: Treat the reasoning slider as a cost-control knob. In Codex, for instance, simple refactoring needs minimal effort; complex algorithm design deserves more.
  5. Admins, lock it down or open it up: Review the ChatGPT admin console to enable or disable Work for your organization. Consider briefing your team on what data can safely be shared with the agent.

Outlook: More Integration and Model Refinement Ahead

OpenAI has signaled that this is just the beginning of a more integrated ecosystem. We can expect cloud syncing for Work threads to arrive eventually, along with more nuanced model routing that automatically picks the right GPT-5.6 variant for a prompt. For Windows users, the app’s ability to interact with local files hints at a future where AI assistants become OS-level tools, not just web services. As the line between chat, code, and agent blurs, the user’s real job will be learning to orchestrate these capabilities—not just type prompts.