Image: Simon Willison
Why it mattersThe cloud version has code execution with internet access, a headless browser and a filesystem that persists between sessions. Those are the pieces of an agent platform, sold as a tab in a chat app.
OpenAI announced ChatGPT Work on 9 July and has been changing it since. Simon Willison published his findings on 30 August, and his opening description is accurate: it is an extraordinarily confusing and very powerful product.
It is two products
Work Cloud runs in OpenAI's cloud, reachable from chatgpt.com, the mobile apps, and the desktop app through a "Where should this chat run?" dropdown.
Work Local comes with the ChatGPT desktop app, the one previously called Codex. It reaches files and runs programs on your own machine. Willison describes it as Codex given a new look to be less intimidating to people who are not software developers.
Both are limited to subscribers paying $20 a month or more. Free and $8 Go users have neither.
What Work has that Chat does not
OpenAI's official guidance is to use Chat for an answer or a short draft and Work for a task with a clear outcome. Willison calls that almost useless, on the grounds that people have been using Chat for all of those categories for years.
His list of the actual feature differences is the useful part:
- Luna and Terra available in place of Sol
- A code execution environment with internet access
- A headless Chrome browser
- A filesystem that persists between sessions
- Publishing ChatGPT Sites
- Sub-agent sessions with Sol, Luna and Terra
- Scheduled prompt automations
Model selection differs too. Work offers GPT-5.6 Sol, Luna or Terra at Light through Ultra, plus GPT-5.5 up to Extra High. Chat offers 5.6 Instant, Medium, High, Extra High and Pro, with Extra High and Pro reserved for subscribers at $100 a month and above. 5.6 Pro appears to be Chat only.
Willison believes Work sessions bill against the Codex allowance while Chat sessions have their own, which would explain the split. He flags that as his understanding rather than documented fact.
The feature that matters
Code execution that can reach the internet.
Chat cannot do this. Ask it to install a package or call an external API and the container proxy blocks it. Willison notes it briefly gained package installation in January and no longer appears to have it, along with a complaint worth repeating: there are no good changelogs for any of this.
For comparison, he points out Claude's container has allowed restricted internet access since last September, including installing from PyPI and npm and cloning from GitHub.
Why a developer should care
Persistent filesystem, headless browser, sub-agents and network-capable code execution are the components of an agent platform. They arrived inside a consumer chat application, behind a tab selector, with no clear announcement of what changed when.
That is the practical problem. A capability that appears without a changelog is a capability your organisation acquires without deciding to. If people on your team have $20 subscriptions, some of them now have a cloud machine that can install software and reach the network, sitting outside anything you administer.
Worth knowing what tab everyone is on.
Source
- Understanding ChatGPT Work, Simon Willison, 30 August 2026
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
