ChatGPT Desktop Overhaul: Chat vs Work Modes, Cloud Sandboxes & 100GB Library

1. Chat Mode vs Work Mode

OpenAI defines distinct product boundaries: ChatGPT for creating, learning, and exploring; Codex for building, debugging, and deploying. The desktop app now offers two modes:

Chat Mode – Core is conversation and communication. It can use GPT-6 Pro without consuming weekly quota, making it ideal for requirement discussions, product design, and iterative questioning. Chat mode also includes a cloud execution environment with search, file analysis, and safe code execution, blurring the line between "chat only talks" and "work does". Projects can be created in Chat mode to group related long-term tasks.

Work Mode – Slogan "What shall we do?" Focuses on concrete, multi-step tasks and complete deliverables. The author provides a decision table mapping scenarios to the recommended mode:

Explain an AI concept / discuss technical approach → Chat – iterate questions to clarify.

Rewrite a public account title, polish paragraphs, refine article angles → Chat – you lead creation, it assists polishing.

Research several models, collect data, verify parameters, deliver report and comparison table → Work – define scope, sources, and output format; let it run the full process.

Check multiple Excel files, fix formulas, merge data, return modified files → Work – execute directly on files, not just advise.

Produce complete Word, PPT, or website from materials → Work – specify final product requirements and review results.

2. Work in the Cloud

Work mode supports both local execution and OpenAI's cloud sandboxes. The cloud environment runs on OpenAI-managed infrastructure using VM-backed sandboxes , sharing part of Codex's execution and isolation mechanisms but with different tools and permissions.

Hardware specs observed by the author:

Plus plan: 4-core CPU, 4 GB RAM Linux container.

Pro plan: 9-core CPU, 9 GB RAM.

Cloud mode suits long-running tasks, scheduled jobs, and any work you prefer not to run locally. A concrete example: while traveling with an unstable connection, the author had the cloud sandbox pull a GitHub repository and debug remotely. Another advantage: tasks continue after you leave the conversation or shut down your computer.

A second table guides environment choice:

Search public materials, research GitHub projects, compile technical reports → Work·Cloud – independent of local environment.

Upload article reading data, compute monthly averages/medians, generate charts → Work·Cloud – process uploaded materials immediately.

Upload Word/Excel/PDF, require organization, modification, delivery → Work·Cloud – no need to touch local originals.

Read email, cloud storage via authorized connections → Work·Cloud – provided the connection supports the needed operations.

Directly organize local Obsidian/iCloud directories → Work·Local – requires direct read/write access to that directory.

Operate Mac apps, use existing local workflows → Work·Local – depends on local apps and permissions.

Develop/debug in existing Python/Swift projects → Codex, choose execution location per project environment – needs code and dev tools.

Benchmark your own GPU / model deployment environment → Your own device or dedicated server – results must reflect the target hardware.

The author notes that sandbox execution state, chat history, Library files, and project files have different lifecycles; a persistent chat does not mean the original process is still running. He recommends treating the environment as a task workspace , not a sole storage location: source code in proper repos, important outputs downloaded or saved to Library/projects, and long-running services (websites, databases, model serving) deployed to proper production environments.

3. Library: The Overlooked OpenAI Cloud Drive

Pro subscribers receive a 100 GB cloud drive called Library. It can also mount other cloud drives via plugins (e.g., Google Drive). The author stores historical Markdown articles and reading data, frequently asking ChatGPT to suggest topics, optimize titles, and run monthly data analyses.

4. Image Mode: Parity with Web

Previously the web version produced better images than Codex; now the desktop image mode matches the web exactly, including drawing modes and templates.

5. Plugins: Growing Ecosystem

The plugin ecosystem (previously covered for Codex) continues to expand. Many plugins are paid or not optimized for Chinese users, but the platform's power is undeniable. The author regularly browses for interesting plugins.

6. ChatGPT Map: Travel Companion That Doesn't Understand Chinese Users

The Map feature appears aimed at restaurant reservations, but official documentation lacks clear usage guides. In the author's testing, map-related queries failed to trigger the capability and often errored out, making it unfriendly for domestic users.

Conclusion: Super App Ambition vs. Local Reality

ChatGPT is clearly pursuing a super-app strategy, consolidating chat, cloud sandboxes, code projects, design, maps, and life services into one interface. However, ecosystem barriers prevent it from becoming the all-in-one life entry for Chinese users; local alternatives like Doubao, Workbuddy, and Qianwen will likely capture the everyday desktop first. For pure productivity and personal workflows, though, this desktop integration is transformative:

Quick communication and ideation stay in Chat mode, leveraging GPT-6 Pro without quota limits.

Long-cycle research, data analysis, and automation go to the cloud sandbox, running even with the computer off.

True complex engineering remains with the deeply integrated Codex mode.

The era of dozens of browser tabs and fear of disconnection is over; the desktop app now acts as an always-ready personal workstation.