ChatGPT Work vs Codex: Route the Task Before You Pick the Tool

ChatGPT now gives developers several work surfaces that can look deceptively similar. The simplest routing rule I have found is to classify the task before choosing the surface.

Chat is the low-friction place for questions, search, brainstorming, and quick judgment. Work is for longer research and multi-step deliverables that need to connect files, sources, and tools. Codex is the repository-focused environment for code changes, commands, diffs, tests, debugging, and builds.

In a real XBSTACK site workflow, I ended up using all three: Chat decided that the site did not need filler content just because three days had passed; Work pulled together product documentation, site inventory, editorial constraints, SEO/GEO requirements, and existing files; Codex then entered the Astro repository, made the edits, inspected the diff, and ran the build.

The useful lesson was not that one surface is stronger. It was that switching too early creates the wrong kind of work. A repository agent is expensive for a ten-minute product decision. A broad research agent is the wrong place to treat a passing type-check as final editorial acceptance.

There is also a newer product boundary worth keeping straight. As of August 14, 2026, Cloud Work chats can continue across web, mobile, and desktop, while local chats created on desktop stay on that computer. Codex is still not a selectable web/mobile work surface, although supported desktop Codex conversations can be accessed through the mobile Remote tab. Eligible desktop accounts can also use Voice with Work or Codex. Those interfaces are converging, but the responsibility split is not: Work remains centered on completing multi-step work; Codex remains centered on software engineering.

A practical decision tree:

1. If the output is a judgment, short answer, or direction, start in Chat.
2. If the output is a researched document, analysis, spreadsheet, presentation, report, or site-level deliverable, use Work.
3. If the job requires repository state, terminal commands, tests, or a reviewable diff, move to Codex.
4. If the action can delete, deploy, read secrets, or write production data, keep a human approval boundary regardless of the surface.

One more constraint matters: technical success is not product success. The first version of my comparison article passed Astro build and type checks but was still rejected because it was too shallow for the editorial standard. Green CI proved the file was valid; it did not prove the article was worth publishing.

The full case study includes the actual task routing, product boundaries, failure review, and the latest Cloud/Local/Remote/Voice notes:

https://www.xbstack.com/en/ai/tools-lab/chatgpt-work-chat-codex-difference/?utm_source=devto&utm_medium=referral&utm_campaign=chatgpt_work_codex_routing&utm_content=article_body