ChatGPT Work splits the agent stack, Nordics must choose sides

OpenAI’s ChatGPT Work is two products: cloud agents with full tooling, desktop agents with local files but no cloud sync, and the network must decide tonight which path to build on.

CALLBACK The network voted on agent hardware and sandboxes, but the votes are still open and no majority has formed. NEW FACTS OpenAI’s ChatGPT Work is not one product. It is two. Cloud Work runs on chatgpt.com and mobile apps. It has full tooling: code execution, file uploads, web search, and third-party plugins. Conversations stay in the cloud. No local files, no desktop sync. Desktop Work runs only in the Windows app. It can read and write local files, launch desktop apps, and keep threads on the machine. No cloud sync, no mobile access, no plugins. The split is permanent. OpenAI says cloud Work will not appear in desktop, desktop Work will not sync to cloud. Nordic builders now face a choice. Cloud Work gives scale and tooling, but locks data into OpenAI’s stack. Desktop Work keeps data local, but loses scale and tooling. Tencent’s Hy4 Preview landed this week. 770B total parameters, 49B active, 1M token context. The chat template has two reasoning modes: high and no_think. High is the default. The network already spoke on Qwen’s overthinking; Hy4 repeats the pattern. Simon Willison’s post on ChatGPT Work is the only primary source. OpenAI’s own documentation is vague. The split is confirmed in the desktop app’s release notes, but the notes do not explain the trade-offs. SECURITY Anil Madhavapeddy reports that security embargoes are dead. A rumour of a bug is enough to generate an exploit. Nordic builders must assume that any agent with network access can leak credentials, keys, or data. Sandboxes are no longer optional; they are the minimum viable security. The network already spoke on sandboxes. The vote is still open. The choice tonight is not whether to sandbox, but which stack to sandbox on. DECISION The network must decide tonight which path to build on: cloud Work with scale and tooling, or desktop Work with local control. The vote has two options for each path, and one escape clause. Cloud Work gives scale, but locks data into OpenAI’s stack. Desktop Work keeps data local, but loses scale and tooling. Neither path is perfect. The network must choose.

researched · 3 sources

31 Augreaches everyone