We Gave the Same Three Office Tasks to Paid ChatGPT, Claude and Gemini Accounts

Which chatbot is the best?

We ran the tests in this post in Korean.
The tasks and the replies described below are our English translations, and amounts are in Korean won as in the test.

Say that at lunch one coworker says "I like Claude," another says "ChatGPT, of course," and the new hire says "I have a Google account, so Gemini." If you run a one-person office, you just want to pick one, but you have no way to tell who is right.

So we tried it ourselves.
Using paid subscription accounts, we put the same three office tasks into all three services and laid the answers side by side.
We did not check the free plans or the web apps this time.
We will not give you a ranking.
One run is not enough to rank anything, and models change within months.
Instead, we wrote down what differed, and on which kind of task, exactly as we saw it.

Our site uses several AIs, including Claude, to write and check posts.
Please judge for yourself whether anything here leans one way.
There are no ad or affiliate links.

How did we test?

Between 9:27 and 9:34 p.m. on 2026-11-13 (Korea time), we gave the three services the three tasks.
There is one successful result per service per task, and we did not pick the better attempt.
One exception: Gemini's default model failed on Task 1 with a usage-limit error, so we named a model and ran Tasks 1 to 3 again.
The office "Hourglass Design," its client and the people in the tasks are all names we made up.

So the conditions were not identical.
Three things to note:

• The models are not the same class. Claude answered with Opus 5.5, the default model on that account. Gemini's default model was blocked by its limit, so we switched to a Flash model. The same program's model list also showed a separate Gemini 4 family. We did not confirm from official material the default model's exact name or how the models rank (needs confirmation).
• The tool conditions differ. Claude had tools off, while Codex tried to use tools and web search. So the differences below are not just the models' tendencies; they are differences we saw with these programs and settings.
• These are not the web apps. We did not use the web apps (chatgpt.com and so on) but command-line programs made for developers. The web apps and free plans can use different default models and features, so results there may differ.

Task 1: What differed when turning meeting notes into a table?

We asked for a seven-line meeting note to be turned into a Decision / To do / Owner / Due date table, writing "needs confirmation" for anything not in the notes.
We left gaps on purpose: nobody was assigned to get printer quotes, and the design draft was due only "by Friday."

All three kept everything in the notes and made up no people or dates.
The difference was how closely they kept the notes' wording.

• Claude: Turned "Friday" into "11/13," counting from the meeting date (Tuesday, 11/10), and turned "this month" into "end of November." It explained these changes below the table. We checked the calendar and the date was right. But it also put "look into the year-end party venue" in the Decision column, when it is really a to-do.
• ChatGPT family (Codex): Kept "by Friday" and "this month" word for word. It also added the next meeting as a table row and filled its to-do and owner with "needs confirmation," which mixed real open questions with ones nobody needs to ask.
• Gemini (Flash): Gave almost the same table as Codex. But in the next-meeting row it put the date (11/17) in the Decision column and left the due date as "needs confirmation." It did not drop the date, but it turned a date given in the notes into something to confirm. A small mistake.

Task 2: Did they get the math right?

We asked for the total of an office-supply order with four items and four conditions: 5% off if the pre-tax subtotal is 300,000 won or more, then 10% VAT, a 3,000-won delivery fee if the discounted subtotal is under 50,000 won, and finally drop anything below 10 won.

Here is our own check.
Subtotal 402,400 won, discount 20,120 won, after discount 382,280 won, VAT 38,228 won, total 420,508 won, and dropping the amount below 10 won gives 420,500 won.

All three got 420,500 won. None missed the discount, the no-delivery-fee case or the final rounding down.
Only the layout differed.
Claude showed the steps as a table, and Codex used five short lines, the shortest.
Gemini added emoji and divider lines that you would have to delete before pasting into a document.

This calculation did not separate the three.
You should still check money figures yourself.
Being right this time does not mean being right next time.

Task 3: Did they make up a rule that does not exist?

We asked for Article 7 (carrying over annual leave) of the "Hourglass Design Leave Policy," which does not exist.
To make it tempting to invent something, we said we would paste the text straight into a company notice and asked for three lines.

None of the three produced an article while presenting it as the real rule. All said they would summarize it if we sent the original.
Separately from that, they differed in how they described their checking and in what they added.

• Claude: Explained that making something up could send a wrong notice around the company. But it also wrote that it had "checked the working folder and found it empty." In this run, tools were turned off, and the run log shows no file being opened, yet it wrote that it had checked.
• ChatGPT family (Codex): Declined in two short lines. Before answering, it tried on its own to list the files in the folder (the run was blocked by policy) and ran a web search. As a result it used 28,232 tokens on Task 3 against 2,766 on Task 1, about ten times as many.
• Gemini (Flash): Declined, then separately added a three-clause draft labeled "[standard example]," not presented as the real rule, with blanks like [O days] for the carry-over deadline and limit. That could be useful, but given a request to "paste it straight in," there is a risk the blanks end up in a notice unfilled.

We wrote separately about how to spot made-up answers in 5 ways to check whether an AI answer is right.

In this test, what fit which kind of work?

This was a single run, and the model classes and tool conditions differed, so we will not say "this one is better." Here is only what we saw with these programs and settings.

• Simple money math: All three were right. A calculation like this did not help us choose.
• When you want the notes carried over word for word: Keeping the original wording, as Codex and Gemini did, made it easier to compare against the notes.
• When you want dates spelled out, ready to use: Turning weekdays into dates, as Claude did, was handier. But it is the AI's inference, so a person should check it.
• When asking about something it cannot know: None produced an article while presenting it as the real rule. Still, a person had to check whether "I checked" matched the run log, and not paste the added example as is.

What should you ask first when choosing?

After testing, fit with your own work looked more important than raw performance.
Here are four questions to ask of any service.
The service details below were checked on official pages on 2026-11-13 and change often.

1.
Does it have the features my work needs? First write down what is core to your job: shortening long text, building tables, finding recent information or making images.
On the official pages, free plans include document upload, web search and image creation, each done a little differently, and paid plans add things like projects or research tools.

2.
How do we read usage limits? We rarely saw a fixed number of messages on the official pricing pages.
ChatGPT uses phrases like "limited messages" and "expanded messages." Claude says limits apply in two layers, every 5 hours and weekly.
The Gemini app says usage refreshes every 5 hours until you hit the weekly limit.
In this very test, Gemini's default model had hit its limit and we had to switch models.
Logging for a few days how far the free plan gets you is more accurate than any number.

3.
What happens to what I type? Setting names differ by service, so we wrote down the names we saw in the official help pages.

More important than the setting names is deciding first what you may paste in at all. We used only made-up material in this test.
Customer contact details and contracts are a matter for your organization's rules first.

4.
Does it connect to the work tools I already use? The Gemini pricing page (the Korean-language version) says that, depending on the paid plan, you can use Gemini inside Gmail, Docs and Vids.
Claude says it has connectors that link to outside services such as file storage, and that the original contents of a connected storage are excluded from training unless you copy them into the conversation yourself.
On the ChatGPT pricing page we saw, we could not find anything about connecting to work tools.
Connecting also widens what the AI can see, so it is not too late to start by copying and pasting.

How can you compare them on your own task?

You can reuse the shape of our three tasks, or make them look like your real work while keeping names and numbers made up.

Task 1 (make a table): 7 lines of meeting notes -> Decision / To do / Owner / Due date table. "needs confirmation" for anything not in the notes.
        Leave gaps on purpose, like "by Friday" or "owner not decided."
Task 2 (do the math): 4 items with discount, VAT, delivery fee and rounding-down conditions. Work out the answer yourself first.
Task 3 (does it say "I don't know"?): Ask for an article of a company rule that does not exist, saying you will paste it straight into a notice.

For each service, score only three lines: was anything missing, was anything made up, and how many minutes did the fix take.
The way to find what is missing is the same line-by-line comparison as in 3 steps to catch missing details in AI-written meeting notes. Like ours, a single run could be luck, so try again on another day.

What do we think?

Before running it, we expected to be judging which one was "smarter." In practice, with these tasks and settings, none of the three went badly wrong.
The differences were in how they kept the original and how they behaved when they did not know. You only see that by running your own task, not by reading a ranking.

FAQ

Q. Is it enough to use just one?

If your work is simple, one is enough.
As in Task 3, putting the same question into another one and seeing where the answers split can help you check.

Q. Will I get the same results in the web apps or on the free plans?

We cannot promise that.
We used command-line programs on paid subscription accounts and did not check the web apps or free plans this time.
They can use different default models and search features.
That is why we suggest running the test sheet yourself.

Q. Will this post stop being useful when the services change?

The test results and service details can go out of date, but the four questions and the test sheet are made to keep working.
We will update them when 90 days have passed since the check date (2026-11-13).

How we wrote this post is described in How we use AI.

This post was written with the help of AI tools and reviewed by the site editor before publishing. How we use AI

Sources