In brief: OpenAI’s new agent product ChatGPT Work connects with common enterprise tools, runs either in an isolated cloud microVM or locally with full computer access, and is set to merge with the regular ChatGPT chat mode by year’s end, according to Greg Brockman.
On July 9, OpenAI released ChatGPT Work, an agent product for knowledge workers that includes three new models in fourteen configurations as well as the merger of the ChatGPT and Codex desktop apps. For CTOs, it’s relevant that, according to OpenAI President Greg Brockman, Work is set to merge with the regular ChatGPT chat mode later this year — making Work a preview of how ChatGPT will be used by default in the future by potentially a billion weekly users.
ChatGPT Work connects with common workplace tools such as Slack, email, Google Drive, calendars, CRM systems and project management tools via hundreds of plugins to produce finished work outputs. Technically, the product is built on the so-called Codex harness — the same infrastructure that also powers models, sub-agents, browser use and multi-hour task processing, though without the developer-facing elements such as Git control or diff views. Each task runs in an isolated cloud microVM: Pro accounts receive 8 CPUs, 20 GB RAM and 64 GB of storage, while Plus accounts get 14 GB RAM. In addition, a managed Chrome service is available, through which the agent operates in the browser via tool calls. Three weeks after launch, Work together with Codex is said to have already surpassed 10 million users, while ChatGPT overall reportedly passed the mark of 1 billion monthly active users in June and 1 billion weekly active users this month.
For decision-makers in IT leadership, the product architecture is crucial: on web and mobile, Work runs exclusively in the cloud, with a task that has been started able to be continued seamlessly across devices — for example, a task can be started in the browser, tracked on a smartphone, and retrieved as a finished document or spreadsheet on a laptop. The desktop app additionally offers a local mode in which the agent works directly on the user’s own machine with full computer access, across files and applications. This local mode is effectively identical to Codex, just without the developer-specific interface. However, local tasks do not appear on web or mobile, and there is currently no way to move a task started locally into the cloud — a limitation that must be taken into account when evaluating workflows and governance requirements.
As output formats, Work delivers interactive sheets, documents and presentations as well as so-called Sites: independently hosted web apps and dashboards that the agent can create, share via URL, and keep up to date autonomously. The product landscape surrounding Work is currently confusing, as it combines elements from ChatGPT, Codex both as an app and as the original cloud agent, the ChatGPT Agent, as well as from Atlas and OpenClaw. For CTOs evaluating governance, data access via plugins, and the handling of cloud versus local execution, it is worth engaging early with permission concepts and the security implications of full computer access in local mode, given the announced merger with the standard chat mode.
Source: www.latent.space · Published August 4, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.