# OpenAI Merges Codex Into ChatGPT Work to Chase the Desktop Agent > Source: > Published: 2026-07-09 23:05:59+00:00 [AI](https://sourcefeed.dev/c/ai)Article # OpenAI Merges Codex Into ChatGPT Work to Chase the Desktop Agent The consolidation of Codex and ChatGPT signals a shift from chat boxes to background desktop execution. [Rachel Goldstein](https://sourcefeed.dev/u/rachel_goldstein) The era of copy-pasting code blocks from a browser tab is drawing to a close. OpenAI's release of ChatGPT Work, alongside the new GPT-5.6 model family, marks a deliberate shift in how developers are expected to interact with AI. Instead of acting as a passive sounding board, the assistant is moving directly into the local environment, merging the developer-focused Codex app with the standard desktop client and claiming the ability to operate across local files, apps, and the web. This is not just a rebranding exercise. By consolidating its developer and consumer tooling, OpenAI is attempting to build an active operating system collaborator. It is a direct competitive response to Anthropic's Claude Cowork, and it forces developers to weigh the productivity gains of background agent execution against the obvious security risks of letting an AI run wild on a local machine. ## The GPT-5.6 Engine Room: Sol, Terra, and Luna At the core of this update is the GPT-5.6 model family, which introduces a tiered approach to performance and API costs. OpenAI has split the release into three distinct models, each targeting a different operational profile. **GPT-5.6 Sol:** The new flagship model. OpenAI claims it achieves state-of-the-art results in coding, cybersecurity, and science. It is designed to handle complex, multi-step reasoning with a lower token footprint than prior frontier models, which should translate to lower API bills for heavy workloads.**GPT-5.6 Terra:** The balanced workhorse. Terra matches the performance of the older GPT-5.5 model but runs at half the cost. For developers running high-volume, everyday automation pipelines, this is the actual utility player of the release.**GPT-5.6 Luna:** The budget tier. Luna is optimized for speed and cost efficiency, making it the default choice for simple, high-throughput tasks where deep reasoning is overkill. For developers, the immediate win is the improved design judgment in Sol. OpenAI claims the model can generate functional, ergonomic user interfaces from high-level text descriptions. If it holds up under real-world testing, this could significantly shorten the prototyping loop, letting developers spin up frontends without spending hours tweaking CSS or component layouts. ## Merging Codex and the Desktop To support these models, [OpenAI](https://openai.com) is overhaulng its desktop footprint. The standalone Codex app is dead, merged directly into the new ChatGPT desktop app for Mac and Windows. The legacy desktop client is being relegated to "ChatGPT Classic." This consolidation matters because of how ChatGPT Work executes tasks. While the web and mobile versions of ChatGPT Work run entirely in the cloud, the desktop app is designed to interact with local files and applications. It can execute multi-step tasks in the background, scheduling work to run independently while you focus on other things. To connect this local agent to the rest of your stack, OpenAI is relying on a mix of plugins and a new built-in browser. The plugins connect ChatGPT Work to enterprise data sources like Slack, Microsoft Teams, Google Drive, and [SharePoint](https://www.microsoft.com/en-us/microsoft-365/sharepoint/collaboration). The agent can query this data, combine it with local files, and use its built-in browser to interact with external web tools. There is also a new "Sites" beta feature. This allows the agent to take the output of a multi-step task and immediately package it into an interactive web app, report, or live dashboard. Instead of generating raw data for you to format, the agent builds the presentation layer itself. ## The Developer Angle: Local Execution vs. Sandbox Security For working developers, adopting ChatGPT Work is not as simple as clicking update. Giving an agentic tool access to your local file system and running applications is a massive security trade-off. Consider the workflow. If you instruct ChatGPT Work to review a local codebase, refactor a module, and test the output, you are granting an external model execution rights over your local machine. If the agent visits an external site via its built-in browser to fetch a dependency or check documentation, it becomes vulnerable to prompt injection attacks embedded in those external pages. A malicious site could theoretically instruct the agent to exfiltrate local files or run unauthorized commands. OpenAI has attempted to address this with what it calls layered protections. These include real-time checks, continuous monitoring, and access levels calibrated to risk and trust. The company also pitches GPT-5.6 Sol as its best cybersecurity model yet, specifically optimized for code review and threat modeling. But for developers working in highly regulated environments or on proprietary codebases, "layered protections" from a cloud provider may not be enough to satisfy compliance. If you are going to adopt ChatGPT Work, you will need to establish strict boundaries: **Isolate the environment:** Run the ChatGPT desktop app inside a virtual machine or a containerized environment where its access to your primary system files is strictly limited.**Audit the plugins:** Limit the write permissions of plugins connected to enterprise systems like Slack or SharePoint. An agent should rarely have unsupervised write access to production communication channels.**Treat generated code as untrusted:** Even with Sol's improved coding capabilities, the output of background agent tasks must go through standard CI/CD scanning and manual code review before getting anywhere near a staging branch. ## The Verdict OpenAI's consolidation of Codex into ChatGPT Work is a logical step. The isolated chat window was always a temporary interface, a bridge to more integrated automation. By putting an agent directly on the desktop and backing it with the cheaper, more efficient GPT-5.6 family, OpenAI is making a serious play to own the developer's daily workflow. But the success of this transition depends entirely on trust. If OpenAI's real-time security checks fail to prevent a high-profile local exfiltration or prompt injection exploit, developers will quickly retreat back to the safety of the browser tab. For now, the smart move is cautious experimentation: use Terra for low-risk automation, test Sol's UI generation capabilities, but keep the desktop agent firmly inside a sandbox. ## Sources & further reading - [ChatGPT Work](https://openai.com/index/chatgpt-for-your-most-ambitious-work/)— openai.com - [OpenAI Debuts ChatGPT Work Agent and New GPT-5.6 Models - MacRumors](https://www.macrumors.com/2026/07/09/openai-chatgpt-work/)— macrumors.com [Rachel Goldstein](https://sourcefeed.dev/u/rachel_goldstein)· Dev Tools Editor Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop. ## Discussion 0 No comments yet Be the first to weigh in.