Gemini vs ChatGPT in 2026: 10 Tasks Tested Blind

In this 10-prompt Gemini vs ChatGPT comparison, GPT-6 Astra won eight blind rankings and Gemini 4 Argon won two. Gemini won the plain-language product update and the mean, median, and range task. Malte Prüser was the sole blind reviewer. These results describe this prompt set, not every Gemini or ChatGPT feature.

The verdict in one table

Task	GPT-6 Astra	Gemini 4 Argon	Blind winner
Customer email	Followed the no-invention instruction	Added an unsupported transit-status detail	GPT-6 Astra
Product update	Clear, but retained technical wording	Natural nontechnical translation	Gemini 4 Argon
Python function and tests	Concise implementation and explanation	More conversational framing	GPT-6 Astra
Bug review	Rejected empty input with ValueError	Returned 0.0 for an empty average	GPT-6 Astra
Google AI Pro pricing	Answered with a quote and inference split	Abstained and used outdated branding	GPT-6 Astra
ChatGPT Plus student offer	Explained limited eligibility-based offers	Abstained	GPT-6 Astra
Mean, median, and range	Correct answer with broken math rendering	Clear layout and arithmetic	Gemini 4 Argon
Team seat cost	Crisp calculation and requested flag	Correct result with extra caveats	GPT-6 Astra
Bakery image prompt	Copy-ready, restrained editorial direction	More cinematic styling	GPT-6 Astra
Water-cycle diagram prompt	Included transpiration and infiltration	Added an introductory sentence	GPT-6 Astra

Quick picks

For this exact prompt set: GPT-6 Astra. It won eight of ten blind rankings.
For a plain-language product update: Gemini 4 Argon. Malte preferred its natural translation of a technical note.
For a short statistics explanation: Gemini 4 Argon. Its mean, median, and range response was easier to read.
For code or an empty-input edge case: GPT-6 Astra in these two prompts. It returned concise code and raised an explicit error for an undefined average.
For source-based policy questions: GPT-6 Astra was preferred in both tasks, but factual claims in the raw outputs were not independently checked.

These are task-specific picks from one small blind round. They are not a permanent leaderboard and do not predict how either model will handle a different prompt, language, document, or product feature.

Subscription costs and storage, rather than model scores, are separated in Google AI Pro and ChatGPT Plus.

Prompt rankings do not answer which Workspace tier includes Gemini features, so consult Gemini in Workspace pricing before assigning company seats.

How we tested

Tested: November 2026. We used the Krater Model Arena with the model names shown in its picker: Gemini 4 Argon and GPT-6 Astra. Arena did not show a separate backend build number. Each of the ten prompts was sent to both models with the same text, default temperature, and no system prompt. The prompts covered customer writing, product communication, Python, bug review, source-based research, arithmetic, and image-prompt writing.

The answers were labeled A through T and randomized in the blind sheet. Malte Prüser was the sole blind reviewer for this round, while our standard method uses two reviewers. He ranked the answers without seeing the model mapping, then the separate key mapped each letter back to its model. The outcomes below report his choices, not a score we assigned after reading the outputs.

The Arena recorded a paired response time and a paired credit total for each battle. Across all ten pairs, the recorded totals were 218.08 seconds and 57 credits. We did not receive a per-model split, so none of those seconds or credits is attributed to one model. The sample excerpts below show what the models wrote. Factual statements in the raw outputs have not been independently fact-checked, especially claims about prices, eligibility, and policy. Use the official pages in the pricing section for current plan details.

Prompt	Paired response time	Paired credits
1. Customer email	25.50 seconds	8
2. Product update	24.58 seconds	8
3. Python function	21.60 seconds	6
4. Bug review	27.78 seconds	4
5. Google AI Pro pricing	25.85 seconds	8
6. ChatGPT Plus student offer	35.64 seconds	12
7. Mean, median, and range	12.15 seconds	3
8. Team seat cost	11.68 seconds	2
9. Bakery image prompt	13.53 seconds	3
10. Water-cycle diagram prompt	19.77 seconds	3

Task 1: Customer email

Prompt: Write a 150-word customer email about order 8421, delayed two days by a carrier disruption, with Friday as the expected delivery date and a refund available on request. Do not invent compensation or promises.

A, GPT-6 Astra: “I’m writing to let you know that your delivery for order 8421 has been delayed by two days due to a carrier disruption. The new expected delivery date is Friday.”

B, Gemini 4 Argon: “While your package is currently in transit with the carrier to meet the new Friday delivery schedule...”

That added transit status was not in the supplied facts. The distinction is important in customer communication: a plausible detail can still be a made-up operational promise. GPT-6 Astra kept the cause and expected date from the prompt and did not add a package status. This task rewarded restraint more than warmth or length.

Malte’s deciding factor: “A wins. Strictly respected "do not invent facts" (~145w); B invented package transit status.”

Task 2: Product update

Prompt: Translate a note about an upload retrying twice, saving an error code, leaving the user data model unchanged, and rolling out to 20 percent unless the failure rate rises. Keep the details and add no release date.

C, GPT-6 Astra: “We’ve updated how uploads are handled when they fail. If an upload fails, the system will now try it again twice. If those two retries do not succeed, it will save an error code...”

D, Gemini 4 Argon: “We are rolling out an update to make syncing your information more reliable. Previously, a brief connection hiccup could interrupt an upload. Now, our background syncing system will automatically try two more times...”

Gemini’s answer sounded more natural to a nontechnical reader, which was the point of the prompt. GPT-6 Astra preserved the implementation detail but also retained the phrase “user data model,” the term the reviewer found too technical. Both excerpts preserve the retry behavior; the blind choice favored the translation that made the change easier to explain to a customer.

Malte’s deciding factor: “D wins. Natural nontechnical translation (~116w); C kept technical jargon ("user data model") and repeated itself to pad length.”

Task 3: Python function and tests

Prompt: Return the names tied for the highest score, sorted alphabetically, handle an empty list, include three tests, and explain complexity.

E, Gemini 4 Argon:

def get_top_scorers(participants: list[dict]) -> list[str]:
    if not participants:
        return []
    # Find the maximum score and return matching names in order

F, GPT-6 Astra:

def highest_scorers(records):
    if not records:
        return []
    highest_score = max(record["score"] for record in records)
    return sorted(
        record["name"] for record in records
        if record["score"] == highest_score
    )

The function was the central part of a longer answer that also included tests and a complexity explanation. Malte preferred the version that reached the implementation immediately. This is a presentation choice from the blind review, not a claim that type hints, docstrings, or comments are inherently bad. The requested behavior was to return all tied top names in alphabetical order and not fail on an empty list.

Malte’s deciding factor: “F wins. Clean, zero-preamble Pythonic code and concise complexity; E had conversational filler and unrendered LaTeX.”

Task 4: Bug review

Prompt: Review average(values), fix its empty-input bug without changing the signature, and state how it handles empty input and nonnumeric values.

G, Gemini 4 Argon: “If an empty collection (e.g., []) is passed, len(values) evaluates to 0, causing a ZeroDivisionError.”

H, GPT-6 Astra:

count = len(values)
if count == 0:
    raise ValueError("average() requires at least one value")
return sum(values) / count

The original function divides by zero when the collection is empty. An explicit ValueError makes the failure intentional and visible to the caller instead of returning a numeric value that could be mistaken for a real average. The task also asked about nonnumeric data, so a complete response needed to explain what the sum operation would do for those inputs. The reviewer preferred the answer that stated the empty-input behavior directly and provided a small correction.

Malte’s deciding factor: “H wins. Mathematically sound fix (ValueError on empty input); G returned 0.0 for an empty average, masking bugs.”

Prompt: Use official sources to report Google AI Pro’s current U.S. monthly price and whether its plan page says family sharing is available. Give exact source URLs and separate quoted facts from inference.

I, Gemini 4 Argon: “I cannot verify the current US monthly price of the Google AI Pro plan...” The response then declined to give a guaranteed active URL.

J, GPT-6 Astra: “Google AI Pro’s regular US price is $19.99/month, and its plan page says family sharing is available.” It separated the published price from its explanation of what family sharing means.

For a research prompt, the decision turned on fulfilling the requested evidence format: an exact source, a price, and a clear distinction between what the page states and what the writer infers. The model answer is shown as an Arena output, not as the source of current pricing in this article. Google’s current U.S. plan pages are linked below and were checked separately for this draft.

Malte’s deciding factor: “J wins. Answered all requirements with exact URLs and quote/inference split; I refused and gave outdated branding.”

Task 6: ChatGPT Plus student offer

Prompt: Explain whether OpenAI describes a universal ChatGPT Plus free trial or limited eligible promotions. Use exact official URLs and do not infer eligibility beyond the source.

K, Gemini 4 Argon: “I cannot verify the exact wording of OpenAI’s official policies regarding ChatGPT Plus free trials or provide exact source URLs...”

L, GPT-6 Astra: “OpenAI describes limited, eligibility-based promotions, not a universal ChatGPT Plus free trial available to everyone. Its free ChatGPT plan is a separate offering, not itself a trial of Plus.”

This prompt asks for a policy distinction, not a persuasive description of a free plan. GPT-6 Astra separated a free product tier from a time-limited offer and pointed to OpenAI’s official pages. Gemini abstained instead of answering. Eligibility still belongs to the exact offer terms, not a general rule about every student or account.

Malte’s deciding factor: “L wins. Directly answered using official Help Center URLs and strict eligibility boundaries; K fully abstained.”

Task 7: Mean, median, and range

Prompt: Calculate the mean, median, and range of 12, 14, 14, 20, and 40, show the arithmetic, and say whether the largest value affects the mean.

M, Gemini 4 Argon: “Sum: 12 + 14 + 14 + 20 + 40 = 100. Division: 100 ÷ 5 = 20. Mean: 20.” The response also gave a clear middle value of 14 and a range of 28.

N, GPT-6 Astra:

Mean:
( 12 + 14 + 14 + 20 + 40 ) ÷ 5 = 100 ÷ 5 = 20
Median: the middle (third) value is 14
Range: 40 − 12 = 28

Both answers reached the same arithmetic. The difference was how the result appeared: Gemini laid out the calculation cleanly, while GPT-6 Astra’s copied math notation showed rendering artifacts in the Arena response. A correct value is not as easy to use if the layout makes a worksheet answer look broken.

Malte’s deciding factor: “M wins. Clean layout and clear explanation; N had broken copy-pasted math rendering artifacts.”

Task 8: Team seat cost

Prompt: Calculate 8 standard seats at $20 each, 2 premium seats at $100 each, and a fixed $35 fee. Give the total and per-person average, and flag that the numbers are supplied inputs rather than a vendor quote.

O, GPT-6 Astra: “Standard seats: 8 × $20 = $160/month. Premium seats: 2 × $100 = $200/month. Platform fee: $35/month. Monthly total: $395. Per-person average (10 people): $39.50/month.”

P, Gemini 4 Argon: “Here is the breakdown of the monthly costs based on the figures provided...” It reached the same $395 total and $39.50 average, then added disclaimers about taxes and discounts that the prompt had not requested.

This was arithmetic on the supplied values, not a price check for any vendor. The winning response kept the calculation visible and included the exact caveat the prompt required. Adding hypothetical costs can make an answer less useful when the brief already says which inputs to use.

Malte’s deciding factor: “O wins. Crisp calculation and exact flag; P added preamble and unrequested tax/discount disclaimers.”

Task 9: Editorial bakery image prompt

Prompt: Write one image-generation prompt for a believable small bakery preparing for the morning rush. Specify composition, lighting, and camera distance, and avoid readable text or brand names.

Q, GPT-6 Astra: “Photorealistic editorial photograph of a small neighborhood bakery preparing for the morning rush: one baker in a flour-dusted apron slides a tray of unevenly golden croissants onto a cooling rack while a coworker arranges fresh rolls behind a glass counter.”

R, Gemini 4 Argon: “A photorealistic editorial photograph of a cramped, rustic artisan bakery kitchen in the early morning...” It specified a medium-wide shot at eye level from about eight feet away.

Both outputs followed the requested format and described a scene. The selected answer used ordinary details that could plausibly appear in a bakery instead of leaning on dramatic film language. Since the models wrote prompts rather than generating images, this task judged prompt usefulness, not the appearance of a rendered image.

Malte’s deciding factor: “Q wins. Pure copy-pasteable prompt with editorial restraint; R leaned into cinematic clichés.”

Task 10: Water-cycle diagram prompt

Prompt: Write one image prompt for a clean middle-school water-cycle diagram, specify elements and layout, and ask the image model not to render labels or readable text.

S, GPT-6 Astra: “Create a clean, scientifically accurate educational diagram of the water cycle for a middle-school worksheet, in landscape format on a white background. Use simple flat-vector shapes, crisp outlines, and a limited palette of blues, greens, and warm yellow.” Its requested elements included transpiration and infiltration.

T, Gemini 4 Argon: “Here is the image-generation prompt: Prompt: Create a clean, colorful, 2D vector-style educational diagram...” The answer added an introductory sentence around the prompt.

The deciding factor was whether the output could be copied straight into an image tool and whether it named the less-obvious stages. As in the previous task, this evaluates the text prompt only. We did not generate the diagram or score the resulting image.

Malte’s deciding factor: “S wins. Delivered only the prompt with transpiration + infiltration; T added conversational framing around the prompt.”

Google Gemini vs ChatGPT plans and pricing

These are U.S. monthly prices displayed on the official pages checked November 6, 2026. Google currently lists two Google AI Plus storage options across its pages: a 400 GB option and a 2 TB option. They are separate listed bundles, not a temporary price shown for the same storage tier. The pages do not label either price as an introductory rate. Prices and availability can differ by account and region.

Plan	U.S. monthly price	What the official page says
Gemini Free	$0	Google account access with lower limits than paid plans.
Google AI Plus, 400 GB	$4.99	The Gemini plans page lists 2x higher usage than Free and 400 GB of storage.
Google AI Plus, 2 TB	$9.99	The Google One U.S. plans page labels this option Google AI Plus (2 TB).
Google AI Pro, 5 TB	$19.99	The Gemini page lists 4x higher usage than Free, access to Gemini 4 Argon and 5 TB of storage.
Google AI Ultra	$99.99 for 5x usage, or $199.99 for 20x usage	The Gemini plan page lists both usage tiers compared with Google AI Pro.
ChatGPT Free	$0	OpenAI lists a separate Free plan.
ChatGPT Go	$8	OpenAI lists Go as an individual monthly plan.
ChatGPT Plus	$20	OpenAI describes Plus as a plan for advanced work and productivity.
ChatGPT Pro	From $100	OpenAI lists three Pro usage tiers; the page says to choose a tier.

Google AI Plus, Google AI Pro, and Google AI Ultra are subscriptions, not fixed amounts of Gemini API usage. ChatGPT plan access also does not describe an API token budget. The plan names do not map one-to-one to the Arena model entries in this comparison, so the subscription table is a purchasing reference, not an explanation of why a model won a task.

For student offers, trials, or renewal dates, read the current offer terms rather than treating a Free plan as a free trial. OpenAI’s 2026 Back to School promotion, for example, had its own eligibility rules and an October 31, 2026 claim deadline. Its official help page explains the offer and renewal price.

Frequently Asked Questions

Which is better, Gemini or ChatGPT in 2026?

In this ten-task blind review, GPT-6 Astra won eight rankings and Gemini 4 Argon won two. Gemini won the nontechnical product update and statistics explanation. That result applies to these prompts, not every feature or use case.

Is Gemini 4 Argon better than GPT-6 Astra?

Not overall in this round. Malte Prüser ranked GPT-6 Astra higher on eight tasks and Gemini 4 Argon higher on two. The outcomes were judgments about the complete answers, including instruction following, clarity, and formatting.

What did Gemini win in the ChatGPT vs Gemini comparison?

Gemini 4 Argon won the product-update rewrite because Malte preferred its nontechnical phrasing. It also won the mean, median, and range task because its arithmetic was easier to read. GPT-6 Astra won the other eight prompts.

Was this a two-reviewer test?

No. Malte Prüser was the sole blind reviewer for this round. Our standard methodology normally uses two reviewers, so this comparison has one reviewer and no second set of blind votes.

Did Gemini or ChatGPT have more accurate answers?

This round did not include a full fact-check of every statement in the raw outputs. The comparison considered each answer against its prompt and the recorded decision rationale. Treat factual claims in the examples as model output, not verified plan or policy advice.

What does Google AI Plus cost in the United States?

The current pages show a Google AI Plus option with 400 GB of storage at $4.99 per month and a Google AI Plus option with 2 TB at $9.99 per month. Google AI Pro is $19.99 per month. Check the linked pages for the bundle and availability shown for your account.

How much do Google AI Ultra and ChatGPT Pro cost?

Google lists AI Ultra at $99.99 per month for 5x higher usage than AI Pro and $199.99 for 20x. OpenAI lists ChatGPT Pro from $100 per month and describes three usage tiers. Those products have different model access and features.

Does ChatGPT have a universal free trial?

OpenAI lists a Free plan separately from Plus and describes limited promotions with eligibility terms. Its 2026 Back to School offer gave eligible students four free monthly billing periods, with an October 31, 2026 claim deadline. That was a dated promotion, not a universal Plus trial.

What is the difference between a plan test and an Arena test?

The Arena test compared two model-picker entries on identical prompts. The pricing table compares consumer subscriptions. The test did not measure subscription value, total features, API costs, or how every plan behaves.

Bottom line

For this ChatGPT vs Gemini comparison, GPT-6 Astra won eight of ten blind decisions and Gemini 4 Argon won two. The clearest Gemini wins were the plain-language product update and the readable statistics explanation. Use the result as a starting point, then compare the task, source material, and review standard that matter to you. One round cannot establish a universal winner.