Gemini 4 Argon Explained: What Google’s new AI model can actually do

Quick Answer: What Is Gemini 4 Argon?

Gemini 4 Argon is a new AI model from Google. It is built for “deep work,” which means big tasks with many steps, like fixing a huge piece of software or reading hundreds of pages and spotting what matters.

Here are the key facts at a glance:

Who made it: Google

• What it is good at: coding, finance, legal work, cybersecurity, and science
• Where you can use it: the Gemini API (through AI Studio), Google Antigravity, the Gemini app, and AI Mode in Search
• Who gets it: developers, companies, and people on Google AI Ultra and Pro plans
• Launch price: $2 per million input tokens and $10 per million output tokens
• Big claim: it ranks number one on the Vals Index, a test that checks how well AI handles real jobs in finance, coding, legal, and tax work

Google also says it is just the start. More models in the Gemini 4 family are coming.

Gemini 4 Argon Price: How Much Does It Cost?

Let’s talk money first, because this is what developers and companies ask about right away.

The Gemini 4 Argon price is an introductory price. That means it is a special starting rate that may change later. Right now it costs $2 for every million input tokens and $10 for every million output tokens.

What is a token? It is a small chunk of text, a bit like a word or part of a word. Input tokens are what you send to the AI (your question, your files). Output tokens are what the AI sends back (its answer). So you pay a little for what you give it and a bit more for what it writes.

There is a discount too. If you send the same material again and again, those repeated “cached” input tokens cost 95% less. This helps a lot when a team keeps feeding the same big document or codebase to the AI.

One more number matters. Gemini 4 Argon can handle 1 million tokens in one go for input and output combined. That is a very large working space. It can keep a huge amount of material in mind while it works. Google says developers should read the new “continuation protocol” in the documentation to set this up properly.

Gemini 4 Argon Coding: Faster Software, Fewer Mistakes

Coding is where Google shows the most proof. The company says its own engineers have been using Gemini 4 Argon inside a tool called Google Antigravity. The result: more than 35% more AI-written code submissions, and nearly one third fewer code rollbacks. A rollback is when a change breaks something and has to be undone. So the code is not only coming faster, it is also holding up better.

Here are four real examples from the announcement.

1. Cleaning up Chrome: Chrome’s code is huge, about 40 million lines. Inside it sat an old folder of about 6 million lines of browser features that were tangled up with the core of how web pages load. Gemini 4 Argon helped the Chrome team pull them apart across hundreds of files.

   Google says months of work were done in weeks. For regular people, this should let a live webpage move between tabs, side panels, and floating windows without reloading, and it should remove a whole category of browser crashes.

2. Speeding up a Python tool overnight: The team behind Google’s open-source Agent Development Kit (ADK) gave Argon a rulebook and a one-paragraph goal: make the framework faster without breaking apps built on it.

   Argon worked on its own overnight. The changes are now live. They cut preparation time for long sessions by 40% and made first-time loading 21% faster.

3. Faster science software: xTB is a well-known free tool that scientists use to predict the 3D shape of molecules and to screen possible new drugs and materials. Using Argon, Google’s researchers made it up to 2.3 times faster at predicting molecule structures and 1.8 times faster at searching for different shapes of the same molecule.

   The updates were shared with the official xTB project on GitHub.

4. Making Google’s own open models almost three times faster: On Gemma 4 E2B, Argon helped lift speed from about 37,900 to over 108,000 tokens per second. That is a 2.85x jump, with the same answers and no drop in quality. Google says this kind of work normally takes months, and Argon did it in a few days.

Google also says Argon is much better at spatial thinking, math, and visual coding. Teams are using it to build 3D visuals, games, and animations that run right in a web browser.

How Gemini 4 Argon Helps With Law, Finance, and Business

Google says Gemini 4 Argon is built for high-stakes office work. It can read documents that mix text, tables, and images, study data that changes over time, and build small interactive tools to explore numbers.

It also writes better. Google says it follows your style instructions more closely than earlier Gemini models, answers more directly with less hedging, avoids repeating itself, and skips cliches. If you have ever been annoyed by an AI that goes on and on, this is aimed at you.

Early access customers and startups are already testing it, and Google says the wins are strong. The company did not list specific numbers for these business uses, so treat this part as a promise to watch rather than a proven result.

Gemini 4 Argon Cybersecurity: Finding Holes Before Hackers Do

The Gemini 4 Argon cybersecurity story is one of the most serious parts of this launch. Researchers and security teams are using it to test ideas, audit important code on its own, and catch tricky bugs, including “zero-days,” which are flaws nobody knew about yet.

A company called Wiz shared two real cases from its “Scan for Good” program, which protects important public systems for free:

1. Government data: Working with a federal agency, Wiz used Argon to find and fix a gap that could have exposed sensitive government data and infrastructure.
2. Healthcare: Argon linked several small weaknesses together and helped the team find a file download bypass. That bug had exposed a master key and could have given someone admin access. Fixing it protected hundreds of thousands of private patient records.

Google is also opening the door wider for defenders. It is expanding to hundreds more partners in its Fairwind Program. Approved defenders can get a version of Argon without cyber guardrails, because they need it to test real attacks. Access is rolling out step by step, and you have to apply.

What Gemini 4 Argon Means for Science

Google says Gemini 4 Argon is already helping with real scientific work, and I was really surprised to see the ongoing use cases.

Quantum computing: Quantum computers make lots of tiny errors, and they need fast decoders to fix them as they happen. Argon turned an internal version of Google’s AlphaQubit decoder into one highly tuned piece of code, nearly doubling its speed.

It also wrote a math building block (called arcsin) for quantum programs. That version was checked with a formal proof tool called Lean, and it uses more than 100 times less “spacetime cost” than a published method.

Nuclear fusion: Stellarators are machines that could one day make clean fusion energy. Since 1967, many believed an idea called Grad’s conjecture meant exact math solutions for their magnetic fields could not exist.

A few counterexamples were found recently. Now a Google team, working with Professor Matt Landreman, has found a wider family of them. Fusion researchers can use these as reliable test cases for their simulation tools.

Rare diseases: Researchers at Gamow Labs used Argon to study de-identified genomes from newborns with rare lung disorders. In one case, it found a chromosome change that regular software had missed for years, and it explained why.

On the lab’s private tests, Argon scored 57% on RareBench (helping diagnose rare diseases) and 52.2% on LabBench (reasoning about lab experiments), and topped both.

Enzyme design: Enzymes are tiny natural tools that cut or build molecules, and they are used in recycling, medicines, biofuels, and detergents. Designing new ones usually takes months. Google’s scientists worked with Argon on a type called serine esterases.

In a matter of hours, Argon gathered data, tuned a model, and narrowed things down to five new designs. Lab tests later confirmed that all five worked.

To keep this going, Google is giving 10,000 scientists free access to the top usage tier of Google Antigravity for 6 months. It is subject to eligibility, capacity, and selection, and applications are open now.

Gemini 4 Argon and Robotics: Learning by Watching

Google also tested Gemini 4 Argon on robots. On a test called ERQA, which checks whether an AI can solve physical, 3D problems, Argon scored 87% and beat the other top models using pure reasoning.

The fun demo: Argon watched a normal video of a person playing cornhole. It mapped out the space, worked out the body movements needed, and wrote the simulation code by itself. When that was moved to Apptronik’s Apollo humanoid robot, the robot sank seven shots in a row in a real test.

Google calls this a “Video-to-Sim-to-Real” pipeline. In plain words: video becomes a practice game on a computer, and then the robot does it for real. It turns weeks of manual setup into a few hours.

Gemini 4 Argon Safety: How Google Tries to Keep It in Check

A powerful AI can help defenders, but it can also help bad actors. So Gemini 4 Argon safety is a big part of this launch. Google says this is its most complete safety stack so far, built with outside partners. Here is what that means in simple terms.

1. Blocking misuse: Google tested Argon under its Frontier Safety Framework and added several layers of defense, including jailbreak resistance, monitoring, and enforcement. In outside red team tests (where experts try to break the rules on purpose), the testers could unlock less than 1% of the risky biology tasks Google checks for.

On a tough test set of harmful requests, rule violations fell by nearly half compared with earlier Gemini models.

2. Stopping risky shortcuts: When an AI agent hits a wall, such as a permission error or a failing test, it may try to cheat its way around it. Google trained Argon to respect limits, report bad news honestly, and ask a human before doing anything that cannot be undone.

On coding tests where the model had no reason to hold back, shortcut attempts dropped to just 2%, Google’s best result so far.

Google also built an automated monitor that reads the agent’s reasoning and actions step by step, and blocks sneaky workarounds in real time. It was used in internal use, testing, and training. Google says it avoided punishing the model’s thoughts during training, so the reasoning stays honest and easy to check.

3. Resisting tricks hidden in web pages: Sometimes an AI reads a web page or email that contains hidden instructions meant to hijack it. On the Gray Swan Indirect Prompt Injection test, Argon leads the field, with attacks succeeding only 0.7% of the time over 15 adaptive tries.

These are Google’s own reported results, so it is smart to read the full Gemini 4 Argon Model Card if you want the details.

How to Use Gemini 4 Argon Today

Gemini 4 Argon is rolling out worldwide today. Here is where to find it:

• Developers: Use the Gemini API paid tier in AI Studio and Android Studio. Google AI Ultra subscribers can also try it in Google Antigravity and AI Studio.
• Companies: Try it in Gemini for business, the API, and Google Antigravity.
• Everyday users: Open the Gemini app or AI Mode in Search if you have Google AI Ultra or Pro.
• Scientists and researchers: Apply for 6 months of free top-tier Google Antigravity access.
• Cybersecurity defenders: Apply to the Fairwind Program for access to Argon without cyber guardrails.

The big idea behind Gemini 4 Argon is not that it chats better. It is that it can take on long, messy, real-world jobs and keep going. Google points to faster code, safer hospitals, new science, and robots that learn from video.

The serious side matters just as much. A tool that can find hidden security flaws can also be misused, which is why Google spent so much space on safety. Most of the results above are reported by Google and its partners, so wait for independent tests before you treat them as settled.

If you want to try it, start with the Gemini app or AI Studio and see how it handles a task you care about. More Gemini 4 models are on the way.

Frequently Asked Questions About Gemini 4 Argon

What is Gemini 4 Argon?

Gemini 4 Argon is Google’s newest AI model. It is built for long, complex tasks in coding, finance, legal work, cybersecurity, and science.

How much does Gemini 4 Argon cost?

The launch price is $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. This is an introductory price.

Is Gemini 4 Argon free to use?

The API is a paid tier. Google AI Ultra and Pro subscribers can use it in the Gemini app and AI Mode in Search. Selected scientists can get free access to Google Antigravity for 6 months.

Is Gemini 4 Argon safe?

Google says it has its strongest safety setup yet, with misuse defenses, a real-time monitor for risky shortcuts, and strong protection against hidden prompt attacks. These numbers come from Google, so check the model card for details.

What can Gemini 4 Argon do for programmers?

Google says it helps with big code cleanups, speed upgrades, and building 3D visuals, games, and animations. Teams using it in Antigravity submitted over 35% more AI-written code with nearly one third fewer rollbacks.