Gemini 4 Argon Review: What It Built and How It Compares to Claude

A 3D village with rivers, stone bridges and snowy mountains, where every tree and every house is drawn by code. A black hole rendered from a single Python file. A Minecraft clone that added a deer nobody asked for. These came out of the first public tests of Gemini 4 Argon, the model Google opened to everyone on 13 October 2026.

The same tests also put Argon up against Claude Opus 5.5 on one website brief. Argon finished in about half the time for less than half the cost. Claude's version worked better.

This review covers what Argon is, what testers built with it, how it scores on independent leaderboards, what it costs before and after its intro price ends on 30 November, and how to try it yourself, step by step, without writing code. Our verdict is at the end.

What Gemini 4 Argon is

Argon is Google's general model for coding, research and writing. Google first gave it only to cybersecurity partners through its Fairwind Program on 30 September 2026, saying it was trained to find, validate and patch software vulnerabilities on its own. It opened to everyone on 13 October.

Here is what Google's model card and launch post say, in plain terms:

• It reads almost anything and writes text. You can give it text, images, video, audio and PDFs. It answers in text only, so it does not make images or audio itself.
• It holds a lot at once. Through the Gemini API it has a shared limit of 1,048,576 tokens for input and output together. A token is a small chunk of text, usually a word or part of one. A normal answer is capped at 65,536 tokens unless a developer turns on long output.
• It thinks before it answers. Developers can set its thinking level to low, medium or high. Higher levels take longer and use more tokens.
• Google claims top scores. Google says Argon was #1 on the Vals Index of finance, coding, legal and tax work at launch, and that hidden instructions planted in content it read took control of it in only 0.7% of attempts on Gray Swan's prompt injection test.

Google's own demos include a pagoda temple built in Blender, a racing game built from photos of a real desk, a 3D chess game with a computer opponent, and a ray-tracing engine written in pure Python. Those are Google's showcase pieces. The independent results are further down.

What testers built with it

The YouTube channel WorldofAI ran some of the first public tests and published them on 13 October in "Gemini 4 Argon Is Out, Fable 5.5 This Week?". Credit for everything in this section goes to them. Here is what they showed.

A 3D village with no downloaded assets

A fly-through of a village with a river, stone bridges and snowy mountains. WorldofAI says it was fully generated by Argon without any external tools or assets. Every tree and every house is drawn by code, which means there were no 3D models or textures to buy or download.

A black hole in one Python file

A render of light bending around a black hole. The clip's own text says there is no game engine, no stock footage and no downloaded textures, only the equation for how light bends near a black hole, solved in Python. It matches the pure-Python ray-tracing demo in Google's launch post.

A Minecraft clone with a deer

A block game called Voxelhearth where you can break blocks and craft tools. Argon also added a deer, which is not in Minecraft itself. WorldofAI says it took about an hour and 20 minutes of work in the command line and about three tries.

An Apple computer museum

This one was someone else's test, which WorldofAI reported. The person asked for an interactive museum of Apple computers without giving it a single image or any data. Argon reportedly found the photos itself, wrote its own tool to clean them up and collected nearly 50 years of Apple specs.

A fair reading of all four: these are showcase builds picked by people who liked the results. They show what Argon can do on a good run, not what it does every time.

The head-to-head with Claude Opus 5.5

WorldofAI gave Argon and Claude Opus 5.5 the same prompt: build a website for a jelly fruit brand.

Argon's site looked more alive, and Opus's worked better. That is one prompt from one tester, so treat it as a first impression. It is also a useful one, because it matches the pattern in the benchmarks below: Argon is strong at visual and creative output, Claude is stronger when the job has to run correctly end to end.

How it scores on the independent leaderboards

Google's launch material has Argon winning most rows. The independent boards, checked on 27 October 2026, are more mixed:

Two notes on the table. The Vals Index changes as new models arrive: Argon was #1 at 68.90% when it launched and is #3 today. And Google's own launch chart, as reproduced by WorldofAI, put Terminal-Bench 4.0 at 57.4% for Argon and 66.4% for Opus 5.5, close to what Vals measured.

Terminal-Bench is the one to watch if you want an AI that builds and fixes things for you. It tests whether a model can finish real jobs in a command line, which is the work behind tools like Claude Code. A gap of more than seven points there is large. On the Vals Index, where Argon leads, the tasks look more like analysis: reading a filing, answering a tax question, reviewing a contract.

Does Argon hallucinate?

WorldofAI reported that people using Argon for real work said it gets things wrong and makes up answers, a problem older Gemini models were known for.

The formal test points the other way. On Artificial Analysis's AA-Omniscience test of factual knowledge, Argon's hallucination rate was 15%, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. Artificial Analysis called it the lowest of any model scoring 45 or more on its index.

That rate needs context. Argon got 50% of the answers right, against 63% for GPT-6 Astra. It says "I don't know" more often instead of guessing. So it guesses less, but it also knows less on that test. The user reports and the benchmark can both be true: a model that refuses to guess on a quiz can still make things up halfway through a long build.

Whichever model you use, check names, numbers and links in anything it writes before it goes to a client.

What Gemini 4 Argon costs

There are two ways to pay: a monthly Google AI subscription for the apps, or per token through the Gemini API.

Subscriptions

Google lists Argon in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers. It is not included in the free Gemini plan or in AI Plus. Engadget reported in May 2026 that AI Pro costs $20 a month in the US and that AI Ultra starts at $100 a month, with a $200 tier above it. Ultra gives more Argon usage. Prices differ by country, so check Google's plans page for yours before you subscribe.

The API, per token

Some details that change the real bill:

• Argon is not on Google's free API tier. You need the paid tier, which means a billing account.
• Output price includes thinking tokens. A high thinking level costs more even if the visible answer is short.
• Cached input, text you send again and again like a long brief, costs 95% less. Argon only does this automatically; developers cannot set up their own cache for it.
• Argon is wordy. Artificial Analysis counted 110 million output tokens to run its index, and output is the expensive side. Its cost to run the index came to $1.62 per task.
• It writes slower than average, at about 81 tokens per second on Artificial Analysis's measurement. It still finished WorldofAI's site in half the time, so speed per token and speed per job are different things.

So the saving in WorldofAI's test came mostly from the intro price and from Argon finishing faster. From December the per-token prices are the same, and the bill will depend on which model uses fewer tokens for your kind of work.

How to try Gemini 4 Argon yourself

There are three routes. The first suits most business owners. The second is for building websites and apps. The third is for anyone who wants to pay per use instead of monthly.

Route 1: the Gemini app (Google AI Pro or Ultra)

1. Go to gemini.google.com or open the Gemini app on your phone, and sign in with the Google account that has your AI Pro or Ultra plan.
2. Open the model menu and choose Argon. If you do not see it, check that the right account is signed in. Google is rolling the model out in stages, so it can take time to reach every account.
3. Type your request and send it. Expect it to take a little longer than a quick chat model, because Argon thinks before it writes and writes at a below-average speed.
4. For a quick research question, use AI Mode in Google Search instead. Google lists Argon there for the same two plans.

Route 2: Google Antigravity, for building pages and apps

Antigravity is what Google calls its agentic development platform: an app, a command-line tool and an editor where an AI agent builds software for you, the job Claude Code does on the Claude side. The app runs on Windows, macOS and Linux, and its developer tier is free, though Argon itself needs a paid Google AI plan.

1. Download it from antigravity.google and install it like any other app. If you already have it, update to the latest version, which is what Google's Antigravity team says you need for Argon.
2. Sign in with your Google account. The Antigravity site lists Argon for AI Pro and Ultra subscribers. Google's main launch post names Ultra only, so if you are on Pro and cannot see it yet, the rollout has not reached you.
3. Choose Argon in the model setting, create an empty folder for the project, and give it your brief (there is one below).
4. Let it work. WorldofAI's full website took about 15 minutes, so do not expect a finished page in seconds.
5. Open the finished HTML file in your browser and test it the way the next section explains.

Route 3: Google AI Studio, pay per use

1. Open Argon in Google AI Studio. This is the link on Google's own model page.
2. Sign in with your Google account. Argon is not on the free tier, so you need a paid API project with billing set up. Google's launch post also says AI Ultra subscribers can try it in AI Studio.
3. Choose a thinking level if the settings show one. Start on medium. Use high only for hard problems, since thinking tokens are billed as output.
4. Paste your prompt and run it. To keep the bill predictable, start with short jobs and check your usage in your billing account before you run anything long.

Run a fair side-by-side test against Claude

Leaderboards tell you how a model does on average. Your own brief tells you how it does on your work. This is the test we would run before moving anything to Argon.

1. Write one brief and freeze it. Use the same words for both models. If you change the prompt halfway, the comparison is worthless.
2. Run it in Argon through any of the routes above, and note the start and end time. On the API, note the cost too.
3. Run it in Claude with Claude Opus 5.5, in the Claude app or in Claude Code, and note the same two numbers.
4. Open both results side by side in your browser, at full width and at phone width.
5. Click every button and link in both. Looks are easy to judge. Broken features only show up when you use the page, and a broken order button is what a customer notices.
6. Ask each model for one round of fixes with the same feedback, and see which one fixes it without breaking something else.
7. Score them on time, cost, looks, working features and how well the fix round went. Pick the winner per job type, not overall.

Here is a starting brief for a website test. It is ours, written for a quick comparison:

CopyBuild a single-page website for a small brand that sells jelly fruit snacks.
Include a hero section with an animated product shot, three flavors with prices,
a short story about the brand, customer reviews and an order button.
Make it bold and colorful, and make every button work.
Put everything in one HTML file I can open in my browser.

And one for an analysis test, closer to what the Vals Index measures. Attach a document you know well, so you can tell when an answer is wrong:

CopyRead the attached document. List the five facts a business owner most needs
to act on, each with the page or section it came from. Then list anything the
document does not say that I might assume it does. If you are not sure about
something, say so instead of guessing.

Argon or Claude: our verdict

Pick Argon when the job is visual and one-off: landing pages, 3D scenes, product demos, a pitch page that needs to look alive. It is half price until the end of November, and testers found its output more colorful and detailed. It also sits near the top of the Vals Index, so it is worth a try on finance and legal analysis.

Pick Claude Opus 5.5 when an agent has to finish real work on its own, in a codebase or a command line. It leads Terminal-Bench 4.0 by more than seven points and tops the Artificial Analysis index. For the agents and automations we build, that is the number that decides it, so we are staying with Claude for that work for now.

Argon is a poor fit if you need images or audio from the model itself, if you are on the free Gemini plan, or if you need a fixed monthly bill after 30 November and have not measured how many tokens your jobs use. In every case, run your own brief before you switch.

If you want help choosing and wiring a model into your own marketing workflows, that is what our AI automation team does.

Frequently asked questions

Is Gemini 4 Argon better than Claude Opus 5.5?

It depends on the job. Argon is #3 on the Vals Index of finance, coding, legal and tax tasks at 69.24%, ahead of Opus 5.5 at 66.97%. Opus 5.5 leads the Artificial Analysis Intelligence Index (58 vs 53) and Terminal-Bench 4.0 (65.15% vs 57.58%), the test of real coding work in a command line.

How much does Gemini 4 Argon cost?

On the Gemini API it costs $2 per million input tokens and $10 per million output tokens until 30 November 2026. From 1 December 2026 it is $4 and $20, the same as Claude Opus 5.5. It is not on Google's free API tier.

Is Gemini 4 Argon free?

No. It is not on the free Gemini plan, not in Google AI Plus, and not on the free tier of the Gemini API. You need a Google AI Pro or Ultra subscription for the apps, or a paid API project billed per token.

How can I use Gemini 4 Argon?

Google lists it in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, in the Google Antigravity desktop app and CLI (the Antigravity site lists Pro and Ultra; Google's launch post names Ultra), and through the paid tier of the Gemini API in Google AI Studio for developers.

When was Gemini 4 Argon released?

Google announced it on 30 September 2026 for cybersecurity partners in its Fairwind Program, and opened it to the public on 13 October 2026.

Does Gemini 4 Argon hallucinate?

Some early users reported made-up answers. On Artificial Analysis's AA-Omniscience test it had a 15% hallucination rate, the lowest of models scoring 45 or more on the index, but it also answered fewer questions correctly (50% vs 63% for GPT-6 Astra) because it says it does not know more often. Check facts in anything it writes.

Sources

• Google: Introducing Gemini 4 Argon (13 Oct 2026)
• Google: Gemini 4 Argon model card
• TechCrunch: Google releases Gemini 4 Argon (30 Sep 2026)
• WorldofAI: Gemini 4 Argon Is Out (YouTube, 13 Oct 2026)
• Vals AI: Gemini 4 Argon and Terminal-Bench 4.0
• Artificial Analysis: Gemini 4 Argon
• The Next Web: Argon's AA-Omniscience results (2 Oct 2026)
• Google Gemini API pricing
• Anthropic pricing
• Google Antigravity: Gemini 4 Argon in Antigravity
• Engadget: The Google AI Ultra plan now starts at $100 a month (19 May 2026)