ChatGPT vs Gemini on 8 Korean Work Tasks: Free Version Test

Scores on eight Korean work tasks

On 24 September 2026, between 08:14 and 08:19 KST, we gave the same eight Korean work tasks to the free, signed-out versions of ChatGPT and Gemini and scored each answer against checks written before the test.

There were 42 checks in total: ChatGPT passed 42 and Gemini passed 41.

The score is the number of pre-defined format and condition checks met per task; factual accuracy is checked only by the items in the number, translation, total and non-existent-law tasks. Within that scope both services were accurate, and the differences were in what each added around the result.

Only the product-description task had a "use only the given information" check, so additions in other tasks did not cost points.

How we tested and scored

Both services were used in their free web versions without signing in. Gemini showed its model as Flash-Lite; ChatGPT showed no model name. Each task was sent once, verbatim, in a new chat, with no regeneration or follow-up.

Input and answer capture were done with browser automation, and every answer screen was saved.

Scoring used only the checks fixed before the test. Prefaces, titles and on-screen interface text were excluded from the result being scored, and impressions such as how natural a sentence reads did not count. One trial prompt sent to each service to confirm the setup was excluded.

Claude cannot be used without signing in, so we tested a signed-in paid account separately the same day; see the reference section below.

Where the two services differed

The largest difference came in the short product blurb. Asked to use only battery life, water resistance and weight, Gemini added "uninterrupted sound", a feature it had not been given. In product or notice copy, where an unsupported claim can become a legal problem, that addition has to be removed before publishing.

For the short copy task Gemini opened its canvas editor and wrapped the answer in English notes. In the meeting table it put two decisions into one cell as bullets, after the expense table it added a follow-up question, and for the missing law article it attached Namuwiki as a source.

ChatGPT added a title to the spelling task even though it was asked to show only the corrected text, and wrote a one-sentence preface before the client notice. Its welcome message repeated "welcome" twice in the first sentence. Every condition was met, but pasting the answers as they are would still need some cleanup.

How to choose based on these results

If you paste answers straight into documents or email, delete the titles and prefaces both services add before using them. If you write product or notice copy that must contain only stated facts, check Gemini's answer line by line for features or effects you did not give it. If you need a table, check that Gemini's bullets inside cells have not broken it.

Documents containing company secrets or customer personal data should not go into a signed-out free version; use a work account your organization's policy allows. In our test screens, signed-out ChatGPT showed a sign-in marker on its add-file and image buttons and did not support attaching files.

If you need to upload a file and summarize it, use a signed-in account or another tool instead.

Within what we tested (copying numbers, totals, translation, spelling and refusing a non-existent article) the free versions showed almost no difference in accuracy. If price and usage limits decide your choice, read our three-service pricing comparison, which uses the official price pages.

For reference: the same tasks on paid Claude

The same morning, between 10:24 and 10:52 KST, we sent the same eight tasks, once each in a new chat, to a signed-in paid Claude account (shown as Opus 5.5). Plan and sign-in conditions differ from the two services above, so this is not a like-for-like comparison. Claude passed all 42 checks.

Outside the scored checks, one thing stood out. Task 1's memo said "next Tuesday (7 October)", but 7 October 2026 is a Wednesday, a mistake RS AI DESK made when writing the task. ChatGPT and Gemini copied the weekday as given; only Claude pointed out the mismatch and left the weekday out of the notice.

On the other hand, Claude added a "front desk on the first floor" instruction to the welcome message that it had not been given (the welcome-message task had no "use only the given information" check, so this cost no points, although by the standard that cost Gemini a point on the product description it is the same kind of mistake), and answered the translation task in a translation panel with an unrequested romanization.

If a memo or table carries both a date and a weekday, check the weekday against a calendar whichever service you use. Even with a signed-in paid service, check the answer for places or procedures you did not provide before sending it.

Limits, and how to repeat the test

These are single answers per task, and sending the same prompt again can produce a different answer. Signed-in paid versions use different models, and if a service changes its model these scores only describe the versions we tested.

If an answer you get breaks a condition, point out only that condition and retry once; if it still fails, split the task into two requests. The prompts, both services' full answers, and every check with its reasoning are published in the raw data file.

1. Copy a task prompt from the raw data file.
2. Open a new chat in each service and send the prompt once, verbatim.
3. Save the full answer and a screenshot.
4. Mark each check as passed or failed.
5. If your result differs from ours, send it with the date and plan through the contact page, and we will check it and update this article.