Understanding ChatGPT Work

OpenAI unveiled ChatGPT Work on July 9th, and it has undergone rapid refinement since then. This is an exceptionally perplexing and highly potent product. The cloud-based version of ChatGPT Work, known as Work Cloud, can be accessed through chatgpt.com or the ChatGPT mobile applications. In contrast, the desktop app—formerly referred to as Codex—provides access to ChatGPT Work, which can interact with files and execute programs directly on your computer.

This latter version is perceived as a more approachable adaptation of Codex for non-developers. Throughout this report, I will exclusively discuss Work Cloud. Presently, ChatGPT Work (in both forms) is exclusively available to subscribers paying $20/month or more. Free users and those on the $8/month Go plan lack access. To determine when to employ Chat and when to utilize Work, OpenAI provides the following guidance: utilize Chat for obtaining answers, explanations, brainstorming, or generating short drafts; resort to Work when seeking ChatGPT to complete a task with a clear outcome, such as a brief, deck, analysis, routine update, workflow, or file that can be reviewed and utilized.

This guidance appears rather inconsequential, as I have consistently employed regular ChatGPT Chat for all of those task categories over the years. The more pertinent inquiry then becomes what features does Work possess that are absent from Chat? After extensive experimentation, I believe I have largely identified these additional capabilities.

In Work, you can select among GPT-5.6 Sol, Luna, or Terra, each offering Light, Medium, High, Extra High, Max, or Ultra reasoning levels. Alternatively, you can opt for GPT-5.5 at Light, Medium, High, or Extra High levels. These appear to be the same models accessible via the OpenAI API. Chat offers a distinct range: 5.6 Instant, Medium, High, Extra High, and Pro (with Extra High and Pro only available for $100/month+ subscribers—$20/month subscribers cap out at High).

It is unclear whether these correspond to Luna or Terra or Sol (my assumption is Sol?). The Pro variant appears exclusive to Chat, with no equivalent in Work. Based on my experience with Codex, Ultra appears to be a specialized mode that delegates tasks more readily to sub-agents. I believe ChatGPT Work sessions consume resources from your Codex allowance, whereas ChatGPT Chat sessions maintain their own separate allowance.

This discrepancy may explain the disparity in model availability. As an ardent advocate of the Code Interpreter pattern, pioneered by OpenAI in 2023, I find the code execution environment in ChatGPT Work (Cloud) to be the most compelling feature. Unlike ChatGPT Chat, which cannot access additional software packages or interact with websites or APIs, Work allows seamless integration with the internet.

While ChatGPT Chat gained the ability to install packages in January, this feature now appears to be non-functional. I wish OpenAI had provided more detailed changelogs. Among the most impressive aspects of ChatGPT Work is its ability to connect to the internet and execute code. ChatGPT Chat lacks this capability. If I request ChatGPT Chat to install additional software packages or interact with websites or APIs, the container proxy will block these actions.

However, Work can clone GitHub repositories, install their dependencies, and leverage them to interact with the broader web. Another standout feature of ChatGPT Work is the browser tool. It can launch a full Chrome instance, load websites, fill out forms, take screenshots, and even run JavaScript against loaded pages. For instance, I prompted ChatGPT Work to load simonwillison.net in its browser and extract headings using JavaScript.

This functionality is reminiscent of my shot-scraper JavaScript tool, but now accessible from my phone. Unlike ChatGPT Chat, which generates a fresh filesystem for each chat session that cannot be accessed by other sessions, Work offers a persistent scratch folder named /workspace/scratch/e00a0a017944 (with variations). These folders persist across sessions, enabling access to files from previous chats.

The /workspace volume is shared among all currently running Work sessions, although they do not share the same process space. Localhost servers running in one session cannot be accessed from another. ChatGPT Work also enables the creation and deployment of entire websites using Cloudflare Workers. These sites can incorporate HTML, JavaScript, and stateful features utilizing Cloudflare D1 and R2.

I constructed a simple site titled london-pelicans-in-her-piety.simonw.chatgpt.site, which identifies locations in London featuring a pelican in her piety—an intriguing element of medieval Christian imagery that can be discovered throughout the city. The site is initially private to the user who created it but can be made public and, on team plans, shared with specific individuals.

Regarding ChatGPT Chat, it cannot run sub-agents, whereas Work can. This functionality is primarily intended for power users working on complex projects that could benefit from multiple parallel agents collaborating.