ChatGPT Work vs Chat: we gave both the same job (and Codex now lives in the same app)

ChatGPT now has a switch at the top: Chat or Work. And since July, the coding tool Codex lives in the same app too. So what is the difference, and when should you use which? We gave both Chat and Work the same job: turn a small data file into a 5-slide presentation. Both got every number right. Chat was about twice as fast. Work showed its plan, made cleaner slides, and its preview worked.

What changed: one app since 9 July

For a long time, ChatGPT and Codex were two separate apps. That changed in the summer. OpenAI's release notes say: "On July 9, the Codex app merged into the ChatGPT desktop app for macOS and Windows."

Codex keeps its own view for programmers. Next to it are Chat and Work. According to OpenAI, the updated desktop app is available worldwide on every ChatGPT plan, including the free one.

Earlier this year, news reports said OpenAI planned one "super app" for all its tools. We did not check those reports. What we did check is OpenAI's own documentation, and it confirms the merged app.

Chat, Work or Codex: what is the difference?

OpenAI's help page explains the three modes in one table. In plain words:

• Chat is for talking things through. Ask a question, brainstorm, write a short message.
• Work is for getting something finished. You describe the result you want, and ChatGPT plans the steps, gathers what it needs and makes a file you can check, such as a presentation, a spreadsheet or a report.
• Codex is for programmers. It shows code, tests and technical details.

Work and Codex share the same usage limits and credits, OpenAI's release notes say. Credits are the units your plan's allowance is counted in. Codex is also the tool we tested in our Codex vs Claude Code article.

Our test: the same job in Chat and in Work

We wanted to see the difference ourselves. So we gave both modes the exact same job in our own ChatGPT Pro account:

We also listed what each slide should show, and asked for the .pptx file. The attached file held the results of our own Codex vs Claude Code test: twelve rows with times and scores. Because we know those numbers, we could check every number on the slides.

We kept the default settings. In Chat that was "Instant". In Work it was GPT-6.1 Sol on the "Light" setting. We did not connect any apps or folders.

What Chat did

Chat was quick. When we checked after about a minute, it was already done: "Done — the 5-slide PowerPoint is ready."

On the way, one step showed "Analysis failed" before it tried again and succeeded. The bigger problem: the preview of the file did not load. ChatGPT showed "Can't load this preview" (in our Dutch app: "Kan dit voorbeeld niet laden"). We tried again later, with the same result.

So we downloaded the file and checked its contents. Every number was right, including a useful extra: an overall average of 159 seconds for Codex against 35 for Claude Code, "4.5× faster". But the slides used our internal task names, such as "a-duration", instead of plain words. We could not see how the slides looked, because the preview failed and we had no PowerPoint program on the test computer.

What Work did

Work took a different route. First it said what it would do: use its presentation skill and calculate the averages from the file. A small panel on the right showed the file it was reading and what it would produce. A timer showed how long it had been busy.

Then it summed up the data before building anything: "Both tools achieved the same secret-test results: 86 of 88 checks passed across their runs." That was correct.

After "Worked for 1 m 58 s", the presentation was ready. Its preview loaded straight away, so we could flip through all five slides inside ChatGPT.

How good were Work's slides?

Good, and correct. We checked every number against our own data:

• Slide 3: a table with 46/48, 30/30 and 10/10 checks per tool, and "86 / 88 checks passed per tool". Correct.
• Slide 4: a bar chart of the average time per task: 76 against 23 seconds, 129 against 40, 272 against 42. Correct.
• Slide 5: a careful conclusion: same test results, shorter times for Claude Code, and a note that this small test "supports a conclusion about these runs, not all coding work".

Work also turned our task codes into plain names, like "Duration" and "Invoice". That made the slides easier to read than Chat's.

Which one should you use?

Based on our test and OpenAI's own advice:

• Use Chat for quick questions, ideas and short texts. It was about twice as fast, and its numbers were right too.
• Use Work when you want a finished file you will actually use or share. It took longer, but it showed its plan, made cleaner slides, and let us check them inside ChatGPT.
• Use Codex if you write software.

Whichever you use: check the numbers. In our test both were right, but OpenAI itself warns under every answer that ChatGPT can make mistakes. If you want an AI that keeps working on its own for days, see our article on OpenAI's dots.

How we checked

On 6 October 2026 we read OpenAI's pages on using ChatGPT, getting started with Work and the release notes. Then we ran the test in the web version of ChatGPT, with the owner's own Pro account and default settings: once in Chat, once in Work, with the same words and the same file. Chat's time is approximate: we checked after about a minute and it was done. Work's time is the one ChatGPT itself showed.

We checked the slides against our own data: Work's in its preview, Chat's by reading the downloaded file. The screenshots are our own. We blurred the sidebar in them, because it shows our private chats and projects. The slides shown were made by ChatGPT Work from our numbers. The charts and diagrams were made by us. Our account is set to Dutch, so some screenshots show Dutch words.

This was one try per mode, with one task. Another try could take a different time or look different. Plans, limits and features change often, so check OpenAI's pages for your own plan.