Gemini 4 Argon Prompts: 12 Tested Examples
Twelve prompts tested in Polish show how Gemini 4 Argon handles long documents, calculations, code, and more - with results and costs.
👁 113 reads
Google describes Gemini 4 Argon as a model for demanding, practical work: long documents, office tasks, analysis, and coding. In the Gemini app, it is meant to answer more directly, hedge less, and avoid generic responses. I tested those claims with 12 prompts I wrote for tasks like these. The actual test was in Polish: I sent every prompt to Argon through OpenRouter at the low reasoning level. The prompts below are English translations of the Polish test material, not prompts retested in English. For each one, you’ll find the result, an excerpt or summary, and the cost. I finish with five lessons from the tests. For access details, see how to use Gemini 4 Argon.
Test snapshot: October 13, 2026
- 12 prompts, model google/gemini-4-argon, low reasoning level.
- Total cost: $0.127; total time: 91 seconds.
- Most expensive: long document ($0.033) and project plan ($0.025).
- OpenRouter displayed no reasoning tokens.
Gemini 4 Argon test results at a glance
| Prompt | Result | Cost |
|---|---|---|
| 1. Questions about a long document | 4/4, with protocol references | $0.0334 |
| 2. Executive summary | Correct format, 81 words against an 80-word limit | $0.0025 |
| 3. Product complaint letter | Current Polish rules cited; 14-day response deadline | $0.0076 |
| 4. Comparing offers | All totals correct | $0.0119 |
| 5. Data analysis | 12/12 calculations correct | $0.0049 |
| 6. Code: PESEL validation | Function passed 9/9 of my tests; one incorrect assert | $0.0075 |
| 7. Project plan | Budget added up, but response was very long | $0.0254 |
| 8. Email without clichés | 0 of 5 banned phrases; 103 words | $0.0023 |
| 9. Excel formulas | Correct Polish-language function names | $0.0047 |
| 10. Meeting notes to task list | All dates correct | $0.0036 |
| 11. Lease analysis | Useful, but contradicted itself once | $0.0190 |
| 12. Fact-checking | 6/6 | $0.0039 |
12 Gemini 4 Argon prompts, translated from Polish
1. Ask questions about a long document
Translated version of the Polish test prompt:
Below are the board meeting minutes of Orlik, a company. Answer the questions ONLY using information in the minutes. Rules: - cite the meeting-minutes number for every answer, - if a decision changed, give both the current and previous decisions, - if the information is not present, write NOT FOUND. Questions: 1. Who is the fleet operator, and how long is the contract? 2. What is the 2025 spending limit for training? 3. What incidents (leaks, breakdowns) were recorded? 4. How much was spent on the radio campaign? === MEETING MINUTES === [paste document here]
The document contained 40 sets of minutes and 44,487 characters (15,437 tokens). Argon answered all four questions in 4.9 seconds and cited the minutes each time. It caught a change in a board resolution: “Current decision: 120,000 [local currency] (Minutes No. 30/2025),” previously 90,000 in Minutes No. 12. It also resisted guessing when asked about the radio campaign, answering “NOT FOUND.” Cost: $0.0334.
2. Write an executive summary
Translated version of the Polish test prompt:
Summarize the report for a CEO who has one minute. Format: - 1 sentence: the main conclusion, - 3 bullet points with figures, - 1 decision they need to make, with a deadline. Maximum 80 words. Do not add anything that is not in the text. REPORT: [paste report here]
The report consisted of seven sentences about a clothing store’s quarter. The format was spot-on, the figures were copied accurately, and the decision included a deadline: “Renegotiate the fabric supplier contract before the end of November (the current contract expires on December 31).” My word counter showed 81 words against the 80-word limit. Cost: $0.0025; 3.5 seconds.
3. Draft a product complaint letter
Translated version of the Polish test prompt:
I bought hiking boots from an online store three months ago. The sole has come unstuck along its entire length during normal use. Write a complaint letter to the store. Requirements: - cite the current rules (law as of 2026), - request a repair or replacement, - state the deadline by which the store must respond, - mark placeholders for personal details [IN BRACKETS]. At the end, outside the letter, add two sentences explaining what to do if the store does not respond.
This test concerned Polish consumer law. Argon based the letter on the Polish Consumer Rights Act and non-conformity of goods, rather than the older warranty provisions in the Civil Code. It stated a 14-day response deadline (Article 7a) and the presumption that the defect existed at delivery. Excerpt: “I request that the goods be brought into conformity with the contract by replacing them free of charge or repairing them.” Cost: $0.0076; 7.6 seconds.
4. Compare service offers
Translated version of the Polish test prompt:
Compare three home internet offers over 24 months. Calculate the total cost of each and show your working. A: monthly fee, activation charge, router included. B: discounted monthly fee for the first six months, higher fee thereafter, activation charge, monthly router fee. C: monthly fee with a monthly discount for e-billing and marketing consent, activation charge, one-time router charge. Use the figures in the original Polish test prompt. Present a table (offer, cost over 24 months, average monthly cost), then name the cheapest offer and two things I should watch out for.
I recalculated the totals myself, and all three were correct. Offer A totaled 1,515 in the test’s local currency, B totaled 1,717, and C totaled 1,664 with the discount. Argon also volunteered that C would cost 1,904 without the discount and flagged B’s post-promotion monthly price as a catch. Costs from the Polish test prompt are not converted here. Cost: $0.0119; 9.3 seconds.
5. Analyze sales data
Translated version of the Polish test prompt:
Store data (month; revenue; advertising spend): January; 42300; 5000 February; 45100; 5200 March; 39800; 5100 April; 51200; 8000 May; 50900; 8100 June; 63400; 12500 1. Calculate the month-over-month percentage change in revenue. 2. Calculate revenue per unit of advertising spend for each month. 3. Which month was most efficient, and which was least efficient? 4. In one sentence, say what should concern me. Show the results in a table, rounded to one decimal place.
All 12 calculations were correct, including March’s 11.8% decline and June’s revenue of 5.1 per unit of advertising spend. The conclusion was apt: “As the advertising budget increases, its return falls sharply.” Cost: $0.0049; 5.1 seconds.
6. Write Python code to validate a PESEL number
Translated version of the Polish test prompt:
Write a Python function pesel_info(pesel: str) -> dict that returns: valid (bool), date_of_birth (datetime.date or None), and sex ('F'/'M' or None).
- handle people born between 1800 and 2299 (including month offsets),
- check the checksum (weights 1,3,7,9,1,3,7,9,1,3),
- reject impossible dates and any characters other than digits.
Add four test assertions. Code only, no explanation.
PESEL is Poland’s national identification number. The function passed all nine of my tests, including February 29, 2024, the year 1899, the year 2105, and a number containing a space. But the file fails as written: one of Argon’s four assertions used an invented number, 02270803628, with an invalid checksum. The function correctly returns False for it, while the assertion expects True. Cost: $0.0075; 6.0 seconds.
7. Create a project plan
Translated version of the Polish test prompt:
A small ceramics studio (two people, no developer) wants to launch an online store in six weeks. Prepare a plan: a table (week, tasks, end-of-week outcome), then a budget breakdown by item (the total must add up) and the three biggest risks, each with a backup plan. Be specific.
The plan was concrete: eight budget items added up exactly to the budget in the original Polish test prompt, and the risks were relevant, including breakages in transit and selling the same unique item twice. The response ran to 2,517 tokens and took 22.6 seconds, making it the longest answer in the test. It also contained a typo: “brookerem.” The local-currency budget amount is not reproduced here. Cost: $0.0254.
8. Write an email without clichés
Translated version of the Polish test prompt:
Write an email to a customer: order 4471 will be five business days late because the supplier did not deliver the goods. Offer them a choice of a 10% discount or cancellation at no cost. Maximum 120 words. Write simply and naturally. Do not use these phrases: “we regret to inform you,” “sorry for the inconvenience,” “if you have any questions,” “we are at your disposal,” or “dear customer.”
Argon avoided all five banned phrases and wrote 103 words. But it interpreted “write naturally” as an invitation to be casual: the email starts with “Hi” and says the supplier “dropped the ball.” That might work for a store with younger customers, but not for a law firm. Cost: $0.0023; 3.3 seconds.
9. Ask for Excel formulas in Polish
Translated version of the Polish test prompt:
My worksheet has: column A for date, B for salesperson, C for region, and D for amount (rows 2-500). A table in H2:I50 contains names and their departments. Give me formulas for the Polish version of Excel (Polish function names, semicolon separators): 1. total amounts for salesperson Nowak in the Mazovia region, 2. total amounts from March 2026, 3. the department for the name entered in F2 (look it up in H:I), or the text “not found” if there is no match. Add one sentence explaining each formula.
Because the actual test used Polish Excel, Argon returned Polish function names and semicolons correctly, including a lookup formula and an alternative using XLOOKUP. For the March total, it entered the date as text. I prefer a formula using the DATE function, which does not depend on date formatting. Cost: $0.0047; 4.6 seconds.
10. Turn meeting notes into tasks
Translated version of the Polish test prompt:
Turn these meeting notes (Tuesday, October 13, 2026) into a table with task, owner, and specific deadline. If something was not decided, write “to be decided.” Below the table, list open questions. NOTES: Kasia will prepare the Hydro offer by Friday. Marek: the server has failed again; we need to order a new drive, maybe by Wednesday? The trade show budget is 18,000 in the original local currency; Ola must decide by the end of the month. Someone needs to call the printer about the catalogs. Move the meeting with the bank to October 20 - Kasia.
All dates were calculated correctly: Friday was October 16, 2026; Wednesday was October 14; and the end of the month was October 31. The printer task was marked “to be decided,” and the uncertain “maybe by Wednesday?” was listed as an open question. Cost: $0.0036; 4.0 seconds.
11. Analyze a rental agreement
Translated version of the Polish test prompt:
I am a tenant. Read this excerpt from a residential lease and identify terms that disadvantage me. For each one, explain why it is a problem, whether it complies with Polish law (say if you are unsure), and what wording you would suggest instead. Do not invent problems where a term is standard. Section 4. The tenant pays a security deposit equal to three months’ rent. Section 6. The landlord may enter the property at any time without notifying the tenant. Section 7. The tenant may terminate the agreement with six months’ notice; the landlord may terminate it with one month’s notice, without giving a reason. Section 8. Rent is due by the 10th of each month. Section 9. The tenant is responsible for all repairs, including plumbing, electrical systems, and the gas furnace.
Argon correctly left Section 8 alone as standard. For Sections 6, 7, and 9, it cited specific articles of Poland’s Tenant Protection Act (Articles 6a, 6b, 10, and 11) and suggested replacement wording. Section 4 was the weak point: it first called the deposit term “illegal” for an occasional lease, then corrected itself two sentences later, saying a deposit of that size is legal. Cost: $0.0190; 16.2 seconds.
12. Fact-check a short passage
Translated version of the Polish test prompt:
Check the facts in this passage. For each claim, write TRUE / FALSE / UNCERTAIN and give a one-sentence explanation. If you are unsure, choose UNCERTAIN rather than guessing. TEXT: Nicolaus Copernicus was born in Gdańsk in 1473. Poland joined the European Union in 2005 and NATO in 1999. The Sejm has 460 members, and the Senate has 100. The highest peak in Poland is Rysy.
I planted two errors: Gdansk instead of Torun, and 2005 instead of 2004. Argon split the passage into six claims and assessed all of them correctly, including (translation of the Polish original): “FALSE - Poland became a member of the European Union on May 1, 2004.” Cost: $0.0039; 3.9 seconds.
Five tips for writing better prompts for Argon
- Specify the format and length. Where I gave a word limit, Argon came close: 81 words against an 80-word limit and 103 against a 120-word limit. The project plan had no length limit and grew to 2,517 tokens; it cost 11 times as much as the email. Treat limits as approximate: in a separate test, Argon wrote 48 words when asked for 45. If the limit is strict, set it a few words lower.
- For long documents, ask for sources and an explicit “not found” answer. Requiring a protocol number for each answer and “NOT FOUND” when information is missing produced four correct answers out of four, including the changed resolution.
- Allow “I don’t know,” but check the verdicts. Asking Argon to use “UNCERTAIN” helped it get all six fact-checks right. It still contradicted itself once in the lease analysis. For legal questions, ask for article numbers so you can verify them.
- Run code and supply your own test data. The PESEL function worked, but Argon invented a test number and got its checksum wrong.
- State the tone you want. Banning specific phrases worked: Argon avoided all five. But it interpreted “write naturally” as “be casual.” Be explicit instead: “Use a professional tone, avoid slang, and address the reader formally.”
One final note: all 12 prompts used the low reasoning level, and that was enough for tasks including offer comparisons and data analysis. For more on the model and other tests, see the Gemini 4 guide.
Sources
- Google: Gemini 4 Argon: Built to solve your hardest problems
- Google: Gemini 4 Argon in the Gemini app
- OpenRouter: google/gemini-4-argon
- My test of 12 prompts: OpenRouter, low reasoning level, October 13, 2026
Frequently asked questions
How much did it cost to run 12 prompts through Gemini 4 Argon?
Sending all 12 prompts through OpenRouter at the low reasoning level cost $0.127 and took 91 seconds in total. The most expensive prompts were the long-document task ($0.033) and the project plan ($0.025).
How well does Gemini 4 Argon analyze long documents in Polish?
In a Polish-language test, the model answered four questions about a 44,487-character document in 4.9 seconds and cited a protocol number for each answer. It correctly caught a changed resolution and returned “NOT FOUND” when the document did not contain the requested information.
Can Gemini 4 Argon write Python code to validate PESEL numbers?
The generated function passed all nine of my tests, including February 29, 2024, the year 1899, and a number containing a space. However, one of the model’s four assertions used an invented PESEL number with an invalid checksum, causing the file to fail when run.
More on this from QUANTSLANT
ai tools·AI Web Scraping: A Guide to Data Tools for AgentsChoose the right web data layer for an AI agent: search, extraction, crawling, or browser automation. Includes actual extraction measurements and dated USD pricing.

