OpenAI merged its coding app into ChatGPT and called the result Work. I gave it two days and a real deadline.

On Thursday morning my Codex app updated itself and came back with a different name. Same dock icon position, same projects, same repositories — but the title bar now said ChatGPT, and next to the Chat and Codex tabs sat a new one called Work, which OpenAI would like to do my job. I keep a spreadsheet of every AI subscription I pay for, what it costs, and when I cancelled it; it is the most honest document I own. So I gave the new agent the most self-referential task I could think of: rebuild that spreadsheet as a live dashboard, watch the prices, and produce a monthly-cost summary I could send my editor without editing it. Then I mostly left it alone for two days.

Normally I'd live with a tool for a full week before writing a word. The news cycle gave me two days, so treat this as a first site inspection rather than a building report. But two days was enough to find the seams, and one of them has a meter attached.

What actually shipped on Wednesday

The announcement, for the record: on July 9, OpenAI folded Codex — the coding agent that has quietly become its most-used serious product — into a single desktop app with ChatGPT, on Mac and Windows, and added ChatGPT Work, an agent that takes an outcome rather than a message. You describe the result you want — a competitor-pricing sheet updated weekly and posted to Slack is OpenAI's own example — and Work gathers context from your connected apps, writes itself a plan, asks you to approve it, and then runs: immediately, on a schedule, or continuously. It comes back with finished spreadsheets, documents, slide decks, and — via a new public-beta feature called Sites — live web dashboards on a shareable URL that keep updating as the underlying data changes. Under it all is GPT-5.6, the model line OpenAI shipped the same day.

Everything is available on every plan, including Free, with the usual tiering: free users get the mid-range Terra model, paid plans can pick between Sol, Terra and Luna. Codex survives as a tab — Sam Altman spent launch day insisting "codex is not going anywhere" and, more tellingly, that "codex is the core of our new work product." Read those two sentences together and you have the whole strategy: the coding agent isn't being retired, it's being promoted.

The merge, from the inside

The update itself was undramatic, which counts for something. My Codex projects, settings and worktrees came through intact, and you can set Codex as the default view so the app opens the way it used to. The old ChatGPT desktop app has been rebranded "ChatGPT Classic," which in software naming is what a hospice bracelet looks like. The standalone Atlas browser is gone too, its agentic-browsing tricks absorbed into the main app and a Chrome extension.

The first thing that broke was mine, technically. I have a small launch script that opens Codex pointed at the right repository every morning; it targets the app's bundle identifier, and the bundle identifier died in the merge — com.openai.codex is now com.openai.chatgpt. Thirty seconds to fix, but if you have real automation built around the Codex app, budget an hour, because every workflow that names the old app is now pointing at a ghost. It is a very 2026 detail that the disruption caused by the autonomous-work launch was to my existing automation.

Why merge at all? The number OpenAI's people kept circling — reported, so hold it loosely — is that of Codex's roughly five million weekly users, about a million were already using it for things that aren't coding: filing expenses, wrangling CSVs, drafting documents in a repo because the repo is where the agent lives. Users had already decided the coding agent was an office worker. Wednesday was OpenAI agreeing with them.

Two days with Work

Setup is the familiar connector dance — I gave it my calendar, Drive, Slack and a folder of billing emails, which is roughly the level of access I'd give a new assistant I intended to watch closely. The plan-approval step is genuinely good: before running, Work shows you the steps it intends to take, and you can strike or edit them. I struck one ("search the web for current promotional pricing" — I wanted my actual bills, not marketing pages) and it took the correction without sulking.

Then it went away and worked, on and off, for the better part of an afternoon. What came back was, honestly, better than I expected. The spreadsheet was correctly structured, with my thirteen current subscriptions, the four I've cancelled this year, monthly and annualised columns, and a change-log tab I hadn't asked for but kept. The slide summary was the right kind of boring. The Sites dashboard is the party trick that might outlive the party: a URL I could open on my phone showing the same data, restyled, with a note that it would refresh as the sheet changed. My editor looked at it for eleven seconds, said "fine," and that is the highest praise a status dashboard can earn.

Now the seam. One line in the sheet said a video tool I subscribe to had raised its price 60 percent this quarter. It hadn't. Work had found the annual plan's price on a billing page and filed it in the monthly column, then — this is the part that gets me — written a sentence in the summary about "notable price pressure in video tooling," a conclusion built entirely on its own filing error. I caught it in seconds, because it's my data and my money. If it had been a competitor-pricing sheet, the kind OpenAI suggests you delegate weekly, I'm not sure I ever would have.

A finished-looking deliverable is a promissory note. Work ships the note beautifully. You still have to audit the arithmetic, and the polish actively discourages you from doing it.

That's the honest trade as of this week. The old chat-assistant failure mode handed you a wrong draft that looked like a draft. The new one hands you a wrong number wearing a chart, inside a deck, on a dashboard, with a URL. The error rate may well be lower; the error's production values are dramatically higher.

The mode problem

The interface has one genuine flaw, and early users on the developer forums found it within hours: Work mode and Codex mode look nearly identical. Both show an agent, a plan, and a stream of activity; nothing on the screen tells you which contract you're operating under. On Friday I typed a repository question into what I believed was Codex and watched Work cheerfully treat my codebase as a business document to be summarised. Nothing broke. But mode confusion in a product whose whole pitch is "it acts on your behalf for hours" is not a cosmetic bug, and "super confusing," the phrase doing the rounds on Hacker News, matches my notes.

The meter

Work tasks are usage-metered. There's no way to say this that doesn't sound like a warning label, so: the continuous and scheduled automations — the very thing that makes Work more than a chat window — bill by what they consume, and a recurring task compounds. A weekly tracker is 52 invoices a year for the same instruction. Anyone who lived through the summer of developers discovering what their coding agents had been spending on flat plans turned metered ones will recognise the shape of this. My rule after two days: schedule nothing you wouldn't be happy to pay for individually, every time, forever, and check the spend page the way you check a parking meter — before the fine, not after.

Meanwhile, on my phone

None of this happened in a vacuum. Two days before OpenAI's event, Anthropic pushed Claude Cowork — its own do-my-actual-job agent — to mobile and the web, in beta, starting with the $100-a-month Max plan. The pitch is the phone as a remote control for work in progress: start a task at your desk, approve a decision from the train, collect the finished thing later, laptop closed. Anthropic also published the number that explains everyone's week: across 1.2 million Cowork sessions, more than 90 percent of usage reportedly isn't coding — it's business operations, content, admin. The same discovery OpenAI made inside Codex, arrived at independently, announced 48 hours apart.

So both labs have converged, the same week, on the same thesis: the chat window is over as the main event, the coding agent was secretly the beta test for everything else, and the product is now finished work. Where they differ is temperament. Work is web-native — happiest reading SaaS tools, filling forms, publishing dashboards to a URL. Cowork is file-native — happiest in your folders and repositories, and its phone story is currently better, because Anthropic built the remote control while OpenAI built the office. I ran both this week; the honest answer is that they're converging on each other's territory at speed, and the moats are about six weeks wide.

Verdict

Scorecard, since I keep one. The merged app stays; the merge cost me one broken script and gave back a genuinely better Codex, and the inline-diff editing and pull-request review in the side panel are the sleeper features of the week for anyone who actually ships code. Work itself: staying on the paid plan for now, with the scheduler used sparingly and every deliverable audited like an expense report.

• Use it now if: your job produces recurring, structured deliverables — status decks, trackers, summaries — from data you already know well enough to spot-check at a glance.
• Wait if: you'd be delegating work whose errors you can't see. Work's failures are quiet, confident and well-formatted, and the worst place for one is a number nobody in the room owns.
• Watch the meter either way. The scheduled automation you forget about is the new gym membership.
• If you live on your phone, Cowork's remote-control model is currently the better week-two experience; if you live in browser tabs and SaaS tools, Work is.

The bigger thing I'll be testing over the next month is the premise itself. Both labs are now betting that you'll hand an agent the boring third of your job and check its output less and less as trust accrues. The tools are good enough that the trust will accrue. Whether it should is the question my one wrong number keeps asking, and it's the question I'd keep asking too — preferably before the dashboard goes to someone who can't tell.