ChatGPT Work vs Codex: Which OpenAI Agent to Use
OpenAI now ships three distinct surfaces, and the rebuilt desktop app puts all of them in one place: Chat for conversation, Work for multi-step business jobs, and Codex for software engineering. They overlap enough to confuse people and differ enough that picking wrong wastes real money.
The short version: Codex is for changing a codebase. Work is for everything else that takes hours.
The actual dividing line
It isn’t “technical vs non-technical,” which is how most comparisons frame it. Plenty of engineers should be using Work, and plenty of non-engineers touch repositories.
The real split is what the agent is operating on:
- Codex operates on a repository. It reads the codebase, writes and edits across files, runs commands, and works within git: branches, diffs, commits, tests. Its whole model of correctness is “does this build and pass.”
- Work operates on your business context. It pulls from connected apps across a 1,400+ app directory, plans a multi-step job, and produces documents, spreadsheets, decks, and research. Its model of correctness is “is this the deliverable I asked for.”
If your task ends in a merged pull request, that’s Codex. If it ends in a file you’d send someone, that’s Work.
Where people pick wrong
Using Work for code changes. Work can write code and it will produce something plausible. What it won’t do is operate inside your repo the way Codex does: branch, test, iterate against failures. You end up copy-pasting and debugging by hand.
Using Codex for research. Codex is anchored to a repository. Asking it to do competitive analysis across your CRM and the open web is using a specialist as a generalist.
Using either for recurring operations. This is the expensive mistake. Both are metered long-run agents, both start only when you or a schedule kick them off. Neither is built for “do this every time X happens”, which is most of what people actually want automated.
Cost shape
Both consume your ChatGPT plan’s allowance rather than carrying separate prices, and both are usage-intensive by design. A long autonomous run is the most expensive thing you can do on a plan. See ChatGPT Work pricing for the tier gates; the same allowance funds both surfaces, so heavy Codex use and heavy Work use compete for the same budget.
Worth knowing: Chat, Work, and Codex are all included on the rebuilt desktop app, so access isn’t the constraint: allowance is.
The same split exists at Anthropic
If you’re comparing across vendors, the mapping is nearly one-to-one: Claude’s Chat / Cowork / Claude Code lines up against OpenAI’s Chat / Work / Codex. Claude Cowork vs Claude Code is the same decision in different branding, and the same trap applies: the coworker product is for business jobs, the code product is for repositories, and neither watches for events.
What neither one does
Both are request-driven. You start a run, or a schedule does. Nothing else can:
- No inbound address: a client, a form, or a teammate can’t hand either of them work
- No event triggers: a new lead, a changed deal stage, an arriving invoice doesn’t wake them
- No unattended write path: approval prompts on write actions mean unsupervised runs stall
For the recurring, event-driven half of the job, Carly is the different shape: cloud agents that fire on triggers rather than requests, each with its own email address so other people can reach them. AI agents start at $35/month, with the Zapier-style workflow steps underneath not metered, which is what makes running something hundreds of times a month viable.
And it composes with both. Carly’s MCP server connects under Plugins in ChatGPT, so a Work run (or Codex) can call Carly’s email, calendar, CRM, workflow, and booking tools directly and actually finish the action.
FAQ
Should I use ChatGPT Work or Codex for writing code?
Codex. It operates inside your repository (editing across files, running commands, working with git), while Work produces code as an artifact you then have to integrate yourself.
Can ChatGPT Work do everything Codex does?
No. Work can write code but doesn’t have Codex’s repository-native workflow. For anything beyond a snippet, the integration overhead makes Work the slower path.
Do Work and Codex share the same usage limits?
Both draw on your ChatGPT plan’s allowance, so heavy use of one reduces headroom for the other. Chat, Work, and Codex are all available on the rebuilt desktop app.
Which OpenAI agent runs automatically?
Neither runs on events. Both start from a request or a schedule you set. There are no webhooks in either surface.
More: What is ChatGPT Work · ChatGPT Work use cases · ChatGPT Work pricing · OpenAI Codex alternatives · Claude Cowork alternatives · Best AI agents for productivity
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
