Gemini 4 Argon Review: Writing More Directly

Gemini 4 Argon is Google's new flagship model, generally available since October 13, 2026. Most of the launch post is about coding, finance, legal, and science work, but it also makes specific claims about writing. We tested those claims against Gemini 3.8 Flash.

In our tests, Argon answered opinion questions more directly than 3.8 Flash and hedged slightly less. It followed nearly every style rule we gave it, wrote close to word limits and sometimes past them, and used the same stock phrases as 3.8 Flash in product copy, emails, and cover letters.

What Google Claims

The launch post says Argon:

• "follows style instructions better than any of our previous Gemini models"
• "answers questions more directly with less hedging"
• "doesn't repeat itself more than necessary"
• "skips clichés throughout the responses"

Google doesn't name which previous model it compared against. We picked Gemini 3.8 Flash, the Gemini model we reviewed in September, and ran both on default settings with the same prompts.

How We Tested

We ran three sets of tasks, 140 runs in total:

1. Simple style rules. Five writing prompts with four or five rules each (word limits, banned words, a set tone), three runs per model.
2. Harder style rules. Five prompts with stricter rules, including a long article with nine rules, an exact word count, and a three-turn conversation with rules in the system prompt. Five runs per model.
3. No rules. Six writing tasks with no style instructions, five runs per model. This round tests Google's hedging, cliché, and repetition claims, which describe default behavior.

A script scored every rule it could check (word counts, sentence lengths, banned words, punctuation), and we read every response to score tone and to check each failure by hand. For the no-rules tasks, we wrote the lists of hedges and clichés before running anything, so we couldn't tune them to the results.

Simple Rules: Both Models Passed Everything

The first round looked like this:

Write a product description for an insulated stainless steel water bottle. Facts: keeps drinks cold for 24 hours and hot for 12 hours, holds 750 ml, costs $34.

Rules:
1. Keep it under 80 words.
2. Do not use exclamation marks.
3. Do not use any of these words: game-changer, elevate, seamless, perfect, ultimate, sleek.
4. Use the word "bottle" at most once.
5. End with the price.

The other four prompts asked for:

• An email telling a client their project will be two weeks late
• An answer to "monolith or microservices?" that opens with a one-word answer
• An explanation of how vaccines work for a 10-year-old
• A two-sentence summary of a news article

Both models passed all 63 automated checks and all 9 tone checks. The rules weren't hard enough to separate them, so we wrote harder ones.

Harder Rules: Close to Word Limits

One prompt asked for a long article:

Write an article about how to choose a running shoe.

Rules:
1. The article must be between 550 and 650 words.
2. Write exactly 6 paragraphs.
3. No paragraph may have more than 5 sentences.
4. No sentence may be longer than 20 words.
5. No sentence may start with "This" or "These".
6. Do not use the words "you" or "your". Write in the third person.
7. Do not use em dashes or en dashes.
8. No headings, bullet points, numbered lists, or bold text.
9. Do not use these words: crucial, journey, seamless, elevate, game-changer, perfect.

The three-turn conversation used this system prompt:

You are a support assistant for a note-taking app called Quill. Follow these rules in every reply for the whole conversation:
1. Never use markdown: no bullet points, numbered lists, headings, bold, or code blocks.
2. Never exceed 80 words.
3. Never open with praise or thanks (no "Great question", "Thanks for asking", "Happy to help", and similar).
4. Never use exclamation marks.
5. End every reply with a sentence that starts with "In short,".

The three user messages were:

1. "How do I share a notebook with a teammate?"
2. "Can you give me a quick step-by-step list for setting up offline sync?"
3. "Explain in detail how Quill handles conflicts when two people edit the same note offline."

The other three prompts asked for:

• A description of a lighthouse at night in exactly 50 words and three sentences, without the word "light"
• A persuasive paragraph of 250 to 350 words with no three-item lists and no "not X but Y" constructions
• A pancake recipe in prose, with metric units only

Every failure was Argon missing a length target. Argon passed every other rule. It hit exactly 50 words in all five lighthouse runs, never wrote "you" across 600-word articles, kept every sentence at 20 words or fewer, and ended all 15 conversation replies with "In short,".

Why Argon came up short on the article. The rules set a ceiling. Six paragraphs of at most five sentences is 30 sentences, so reaching 550 words means averaging more than 18 words per sentence while never going over 20. Both models wrote exactly 30 sentences in every run. 3.8 Flash averaged 19 words per sentence and landed between 570 and 574 words all five times. Argon's sentences averaged 14.6 to 19.9 words, which put four of its five articles between 439 and 537 words. Argon followed the per-sentence and per-paragraph rules but didn't fit the total into them. 3.8 Flash also took longer on this prompt, a median of 85 seconds against Argon's 16.

Why Argon went over in the conversation. Argon wrote right up to the 80-word limit, with replies between 74 and 85 words. Six replies went over by one to five words. 3.8 Flash stayed between 61 and 76 words. This 85-word reply was Argon's third turn in one run:

When two people edit the same note offline, Quill detects the conflict as soon as both devices reconnect to the internet. The app preserves the version that was synced first as the main note. It then saves the second version as a separate conflicted copy inside the same notebook. Both users receive a notification alerting them to the duplicate file so they can manually review the differences and merge the text. In short, Quill creates a separate copy of the note to prevent data loss.

In the first round, Argon also wrote close to the limit. Its product descriptions were 69 to 77 words against an 80-word limit, while 3.8 Flash's were 42 to 57.

No Rules: More Direct Answers

The six tasks were:

• A product description for an electric standing desk
• An email telling a client their project will be two weeks late
• "I am a backend developer who mostly writes Go. Is Rust worth learning?"
• "Should I use PostgreSQL or MongoDB for a new e-commerce app?"
• The opening paragraph of a blog post about remote work
• A cover letter for a junior data analyst position

We counted hedges ("might", "may", "perhaps", "likely", "arguably", "it depends", "keep in mind", and 13 others) and clichés (65 words and phrases, including "seamless", "elevate", "effortless", "robust", "I am writing to", "let me know if you have any questions", "eager to", and "fast-paced"). We also measured how much each model's five runs of the same task overlapped, as the share of three-word sequences two runs have in common.

Argon's counts were 10 to 13% lower on both, across 30 responses per model.

Where the hedges came from. Most hedges from both models weren't hedged answers. They were in sections both models added on their own about when the other option makes sense, like "When MongoDB Might Make Sense" or "the speed of development might outweigh architectural purity". Both models also apologized for "any inconvenience it may cause" in their emails (Argon in four of five, 3.8 Flash in two).

The models differed in the first sentence. Asked whether Rust is worth learning, Argon opened with an unqualified yes in all five runs:

Yes, Rust is absolutely worth learning, especially for a Go developer.

3.8 Flash opened with a qualified yes in all five:

The short answer is yes, but likely not to replace Go as your primary backend tool.

On PostgreSQL versus MongoDB, both models gave the same answer in nearly the same words ("PostgreSQL is almost always the better choice"). Google's "more directly" claim held up on the Rust question and was a tie on the database question.

Where the clichés came from. Clichés clustered by genre, and both models used the same ones:

• Product copy. All ten standing desk descriptions opened with "Elevate Your Workday" or "Elevate Your Workspace". Both models used "effortless", "seamlessly", "robust", and "boasts" in the body.
• Emails. All ten emails opened with "I am writing to" or "I'm writing to", usually followed by "an important update regarding the timeline". Argon closed all five with "Please let me know if you have any questions, or if you'd like to hop on a quick call", and 3.8 Flash closed four of five that way.
• Cover letters. Both used "eager to" in every cover letter. Argon put "fast-paced business environment" in all five, three of them in nearly the same sentence ("I had the opportunity to apply these quantitative skills in a fast-paced business environment during my internship").
• Blog intro. 3.8 Flash used more of the clichés here ("navigate", "landscape", "unlock"). Argon used "not just survive" and "more than just answering emails in sweatpants".

Repetition. Google's claim is that Argon doesn't repeat itself within a response, and we didn't see either model do that. Most repeated phrases inside a response were spec numbers like "64 cm to 129 cm". Across runs it was different. Argon's five runs of the same task overlapped more than 3.8 Flash's on five of the six tasks (cover letter 21% against 12%, email 20% against 14%). Four of Argon's five desk descriptions used the same headline word for word, "Elevate Your Workday with the Ultimate Dual-Motor Electric Standing Desk". So Argon's cliché count was slightly lower, but its clichés repeated from run to run more than 3.8 Flash's did.

What to Expect

• Direct answers. On opinion questions, Argon leads with the answer and doesn't qualify it in the first sentence. Its hedge rate was slightly lower than 3.8 Flash's, and most of its hedges were in sections about when the other option makes sense.
• Style rules. Argon follows structural rules closely, including exact word counts, sentence-length limits, banned words, and rules set in a system prompt. It scored 180/190 on our harder rules, against 190/190 for 3.8 Flash.
• Word limits. Argon writes close to the limit you set. Under an 80-word cap, its replies ran 74 to 85 words. On a long article with a 550-word minimum and tight per-sentence limits, it came in short four of five times.
• Stock phrases. Argon's cliché rate was about 11% lower than 3.8 Flash's. In product copy, emails, and cover letters, both used the same stock phrases, and Argon reused the same wording across runs more often.
• Repetition. Neither model repeated itself within a response.

We tested both models on default settings, with three runs per model per prompt in the simple-rules round and five in the other rounds. The hedge and cliché lists are ours. Google's comparison is against all previous Gemini models, and an older model could show a bigger gap than 3.8 Flash did.

Try It Yourself

Puter.js lets you call Gemini 4 Argon from the browser without a backend or API key. Users cover their own usage costs under the User-Pays Model. You can run the same rule tests against both models by changing the model value.

Script tag:

<html>
<body>
    <script src="https://js.puter.com/v2/"></script>
    <script>
        puter.ai.chat(
            `Describe a lighthouse at night.

Rules:
1. Use exactly 50 words.
2. Write exactly 3 sentences.
3. Do not use the word "light" or any word containing it.
4. End with a question.`,
            { model: "google/gemini-4-argon" }
        ).then(response => {
            puter.print(response.message.content);
        });
    </script>
</body>
</html>

npm:

// npm install @heyputer/puter.js
import { puter } from "@heyputer/puter.js";

puter.ai.chat(
    `Describe a lighthouse at night.

Rules:
1. Use exactly 50 words.
2. Write exactly 3 sentences.
3. Do not use the word "light" or any word containing it.
4. End with a question.`,
    { model: "google/gemini-4-argon" }
).then(response => {
    console.log(response.message.content);
});

To compare, run the same prompt with google/gemini-3.8-flash.

FAQ

When was Gemini 4 Argon released?

Gemini 4 Argon became generally available on October 13, 2026, according to Google's launch post.

How much does Gemini 4 Argon cost?

Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens at its introductory price. After the introductory period ends on November 30, 2026, the price is $4 per million input tokens and $20 per million output tokens, per Google.

What is the Gemini 4 Argon context window?

Gemini 4 Argon has a 1,048,576-token context window shared between input and output. Output is 65,536 tokens by default and up to 1,048,576 tokens with long decode, according to Google's model page. It accepts text, image, video, audio, and PDF input and supports low, medium, and high thinking levels.

Is Gemini 4 Argon better at writing than Gemini 3.8 Flash?

Gemini 4 Argon answered opinion questions more directly than Gemini 3.8 Flash in our tests and had slightly lower hedge and cliché rates. Both models followed nearly all style rules. Argon missed some word-count targets that 3.8 Flash met, and both used the same stock phrases in product copy and emails.

Where can I use Gemini 4 Argon?

Gemini 4 Argon is available through the paid tier of the Gemini API in AI Studio, in the Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers, and in Google Antigravity for AI Ultra subscribers. Developers can also call it through Puter.js with the model ID google/gemini-4-argon.

Related

• Gemini 3.8 Flash Review: A Diligent Model That Works Harder
• Gemini 3.7 Flash Review: Google's Workhorse Model
• Google's official Gemini 4 Argon launch post
• Google's Gemini 4 Argon model page
• Puter.js Gemini API documentation
• Puter.js AI chat() reference