On July 9th, OpenAI released ChatGPT Work, an agent product for knowledge work, which has already attracted over 10 million users. This analysis explores its design, positioning, and implications for ChatGPT's future as it aims to serve a billion users.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
OpenAI released ChatGPT Work on July 9, and within three weeks, combined users with Codex reportedly surpassed 10 million. Based on external reconstruction, this article analyzes its design for memory, persistence, local and cloud modes, and highlights key differences from open-source agents like OpenClaw.
On July 9, 2026, OpenAI released ChatGPT Work, an agent product for knowledge work. The launch was massive: three new models, fourteen configurations, integrating the ChatGPT and Codex desktop apps, and bringing cloud agents to the mainstream.
Three weeks after launch, combined Work and Codex users reportedly exceeded 10 million. ChatGPT is expected to reach 1 billion monthly active users in June and 1 billion weekly active users this month. Work and Chat are currently separate modes within ChatGPT, but Greg Brockman confirmed they will merge before the end of the year.
At its core, Work is a knowledge work agent that can connect to hundreds of plugins like Slack, email, Drive, calendar, and CRM, gathering context across platforms and producing finished deliverables. It runs on the Codex harness, inheriting the same models, sub-agents, browser use, and long-running task capabilities, but the UI hides code traces.
Work runs on isolated micro-VMs: Pro accounts get 8 CPUs, 20GB RAM, and 64GB disk; Plus gets 14GB RAM. It also manages a Chrome service for the agent to operate. Each new conversation is called a 'task,' runs in the cloud, and can be switched across devices.
The desktop app has two modes: cloud mode and local mode. Local mode operates directly on the user's machine, similar to Codex but with code UI removed; local tasks do not appear on web or mobile and cannot be migrated to the cloud.
Work's cloud computer is persistent, but not a single always-on VM. The workspace syncs to persistent storage and is restored to an isolated micro-VM when needed, so the underlying machine is variable but the work state persists. Each task has a working directory under /workspace/scratch, where the agent can freely create folders, install dependencies, run scripts, and maintain databases.
Context between tasks is not directly shared via the working directory. By default, each new thread receives a compressed summary of recent tasks and files; raw conversation history is not stored on the computer and is queried via the Personal Context tool when needed. Files are stored in ChatGPT's Library; user uploads automatically enter it, and files created by the agent are saved when the user requests or when judged valuable.
Uploaded files exist in duplicate: a working copy in the thread and a canonical version in the Library, which are not synchronized. If thread A uploads a file and thread B modifies the Library version, thread A will still read the old copy when resumed. Accessing the scratch directory across tasks requires explicit instructions, and directory names are opaque with no retention guarantee.
Memory is managed externally. ChatGPT's core memory primitive is a continuously synthesized user profile, maintained asynchronously by the product and provided to Work at task start. The agent can reason based on the profile but cannot modify or create autonomous memory like OpenClaw's.
The key difference between OpenClaw and chatbots is that it owns its own computer, can run persistently, install software, maintain databases, and its state spans the entire computer. Work's cloud computer is also persistent, but the agent's sovereignty over the computer is far less than OpenClaw's.
This article is based on external reconstruction by Latent Space, verified through hands-on testing and conversations with Codex, but not officially confirmed by OpenAI. User numbers are 'reported,' and the merger plan is confirmed by Brockman.
ChatGPT Work is a key step for OpenAI in bringing agents to the masses. Its design balances persistence, memory, and user control, but has limitations such as file desynchronization and restricted cross-task access.
Primary report
Primary source