Latent Space Teardown of ChatGPT Work: Pro Agents Get 8 CPUs / 20GB RAM / 64GB Disk, a Separately Hosted Chrome, and 1,000+ Plugins
Latent Space (Shlok Khemani)·high signal
Shlok Khemani's guest analysis (Latent Space, August 4, 2026) reverse-engineers the ChatGPT Work architecture that OpenAI shipped July 9: it runs on the Codex harness with three primitives — a persistent cloud computer (Pro: 8 CPUs, 20GB RAM, 64GB disk; Plus: 14GB RAM), browser control via a separately hosted Chrome service driven through tool calls, and an integrated ecosystem of over 1,000 plugins. Work and Codex crossed 10 million combined users within three weeks, against ChatGPT's ~1B MAU in June and ~1B weekly actives now. Khemani's thesis is that continuity across tasks "runs through an opinionated ChatGPT product layer rather than the computer itself" — the deliberate opposite of unrestricted harnesses like OpenClaw.