On July 9, OpenAI launched GPT-5.6, a family of AI models, and with it a redesigned desktop application for Windows that can reach into your local files, operate applications, and browse the web on your behalf. Called ChatGPT Work, the agent marks the first time the chatbot can autonomously gather context, draft documents, and manipulate spreadsheets and presentations directly on your PC.
Three new model tiers—Sol, Terra, and Luna—replace the old numbering treadmill, but the more consequential change is what ChatGPT can now do with your permission. While you remain the gatekeeper, Work can plan a job, execute it across several tools, and present the result for your approval. It’s a shift from a conversational window to a genuine workspace.
That shift comes with fresh governance questions, new competitive stakes, and an unusual disclosure: OpenAI gave the U.S. government an advance look at GPT-5.6’s cybersecurity capabilities before the public launch.
What actually changed
GPT-5.6 introduces a naming scheme intended to outlast a single model. Sol is the flagship for difficult coding, scientific, and security tasks. Terra balances capability, speed, and cost for everyday use. Luna targets fast, high-volume workloads where cost matters more than peak reasoning. API pricing reflects the tiering: Sol costs $5 per million input tokens and $30 for output, Terra half that, and Luna $1 in, $6 out.
Inside the ChatGPT interface, however, these tiers aren’t all equally accessible. GPT-5.3 Instant remains the default for ordinary conversations. GPT-5.6 Sol powers eligible reasoning options, while Terra and Luna are surfaced primarily within ChatGPT Work and the revamped Codex environment, depending on your subscription plan.
The redesigned desktop application—available now for Windows and macOS—combines three environments: Chat for conventional conversations, Work for extended agentic tasks, and Codex for software development. Existing Windows users may be prompted to install the new version, while the previous client can remain as “ChatGPT Classic.”
ChatGPT Work: outcomes, not prompts
Work is built around deliverables. You can feed it files, describe a desired report or presentation, connect relevant cloud services, and let the agent gather information, devise a plan, and construct the final product. You monitor progress, answer questions, redirect tasks, and approve sensitive actions.
On Windows, Work can access local files and supported desktop applications after you grant permission. Its built-in browser can navigate sites and web tools. Scheduled tasks can run once, at intervals, or in response to changes. This puts OpenAI in more direct competition with Microsoft Copilot and automation platforms that already sit close to business workflows. The difference: OpenAI aims to combine research, document creation, browser operation, coding, and recurring automation in one client.
Codex becomes a first-class citizen
Developers gain an integrated Codex that edits Markdown and source code inline, displays diffs, responds to annotations, reviews GitHub pull requests, and works across multiple repositories. Existing Codex projects and settings carry forward. The model can also visually inspect rendered interfaces—a step beyond generating plausible code without checking for clipping or inconsistency.
A new “ultra” setting, available to Pro and Enterprise customers in Work and to Plus and higher plans in Codex, assigns parallel sub-agents to complex jobs and then combines their output. This can accelerate tasks like modernizing a large application or reviewing several repositories, but it also multiplies the risk of errors propagating across workstreams.
What it means for you
For home users and non-specialists
If you use free or Go plans, you’ll default to the mid-tier Terra inside ChatGPT Work. The most you’ll likely do is draft a letter or format a spreadsheet. But the agent can now read your Documents folder and launch Excel. That’s a leap in integration—and in what you’ll need to trust. Review which folders and services you allow Work to see, because once permission is granted, the agent can act across them without asking for every step.
For power users and knowledge workers
Work becomes a force multiplier when juggling reports, presentations, and multi-step research. You can retrieve data from a local database, summarize it, feed it into a Word template, and schedule weekly updates—all from a single client. The catch: you’ll need to learn its approval boundaries. Let it draft freely, but review formula-heavy spreadsheets and data pulls before they go anywhere.
For IT administrators
This is where the governance alarm sounds. Giving an agent permission to read local folders, connect authenticated cloud services, and act inside desktop applications is fundamentally different from letting employees paste text into a chatbot. You’ll need to review plugin access, data-handling policies, account permissions, and retention settings. The consolidation of capabilities in one client also raises the stakes: a misconfigured agent could inadvertently alter a shared file or leak sensitive content through a browsing session.
OpenAI says Work is available on desktop across all plans, with web and mobile rollout initially targeting Pro, Pro Lite, Enterprise, and Edu, then Plus and Business. Free and Go users don’t get the web/mobile agent yet. Plan accordingly: if your organization relies on web-only access, you have more time to prepare.
For developers
Codex’s tighter integration means you can now code, review, and iterate without leaving the ChatGPT app. The visual inspection capability reduces the guesswork in UI generation. But treat all generated code as untrusted until builds, tests, static analysis, and human review pass. The ultra setting’s multi-agent execution can shorten sprints, but it also increases API call complexity—monitor usage if you’re paying for tokens.
How we got here
OpenAI’s trajectory toward agentic behavior has been public for months. GPT-5.4 and 5.5 already offered reasoning and tool-use improvements, but Work represents a concrete packaging of those pieces. Meanwhile, Anthropic’s Claude Fable 5 has earned praise for coding and agent performance, pushing OpenAI to respond.
The government preview is a new chapter. On June 26, OpenAI disclosed that it had shared GPT-5.6’s capabilities and launch plans with the U.S. government under an executive order framework. At the government’s request, access was initially limited to trusted partners while cybersecurity safeguards were reviewed. OpenAI objected to making that arrangement permanent, arguing it could delay useful tools for defenders. General availability followed on July 9.
GPT-5.6 is OpenAI’s most capable cybersecurity model to date. In browser evaluations, Sol identified bugs and exploit components but did not autonomously construct a full end-to-end exploit. The company says the model remains below its “Cyber Critical” threshold, though improved vulnerability research prompted the additional safeguards. Security teams now face a dual-use reality: faster code auditing and patch development, but also greater potential for misuse.
What to do now
- Install or update the desktop app. If you’re on Windows, check for the new ChatGPT client. The old one can stay as Classic, but Work requires the new app.
- Review permissions immediately. Before using Work, go to Settings → Data controls and see what is connected. Turn off anything you don’t need. Limit local folder access to a dedicated workspace if possible.
- Test with low-stakes tasks. Start by asking Work to create a simple document in a sandboxed folder. Observe how it requests permissions and whether it behaves as expected.
- If you’re an admin, begin governance planning. Inventory which users are on paid plans and could soon access Work on web/mobile. Draft internal policies on data sharing, agent-to-agent actions, and approval workflows. Check whether your existing DLP and logging tools can monitor actions taken by the ChatGPT agent.
- Watch for GPT-5.4 deprecation—but don’t panic. Swiss outlet blue News reports that GPT-5.4 will be shut down on July 23, but this date is not confirmed by OpenAI. Still, you should test any pinned prompts, custom GPTs, or code-generation pipelines against GPT-5.6 now. Changing the model can alter output even when instructions remain the same.
- Developers: migrate Codex workflows. Export your projects from the old Codex environment if needed, and test the new ultra setting for non-critical tasks first. Monitor token usage closely.
Outlook
OpenAI is turning ChatGPT into a platform that can act on your behalf, not just talk about it. The immediate battle is with Anthropic’s Claude, but the deeper one is with Microsoft’s Copilot ecosystem, which already has deep hooks into Windows and Office. The speed of iteration means new models and agent capabilities could arrive within weeks, making the governance challenge a moving target.
The government’s review role is also likely to expand, potentially adding delays or transparency requirements for future frontier models. For Windows users, the choice is now not just which AI to chat with, but how much of your desktop you’re willing to hand over.