Gemini icon via Wikimedia Commons
What Gemini 4 Argon is
Gemini 4 Argon is Google's latest frontier model, and Google announced it as globally available from October 13, 2026. A frontier model is the most capable tier a company offers at a given moment, the one it points to when it wants to show what its research can do. Google describes Argon as just the beginning of the Gemini 4 family, which tells you more models are planned under the same name.
If you have used the Gemini app or seen Google's AI answers at the top of search results, Argon is the newer engine that can sit behind those tools. You do not need to know how it works inside to use it. What matters for most people is which product gives access, what it costs, and whether the claims Google makes hold up for the kind of work you do.
This article walks through all of that in plain terms. Every number below comes from Google's own announcement and its subscription page, and where Google reports a test result, we say so, because those tests were run and described by the company that built the model.
What Google says it can do
Google's post groups Argon's abilities into a few areas. The first is professional knowledge work. Google reports that Argon ranks number one on the Vals Index, a benchmark that covers finance, coding, legal, and tax tasks. That is a Google-reported result, and it is worth reading as a signal about the types of work Google tuned the model for. It says little about how the model will handle your own documents.
The second area is reasoning about the physical world. Google reports 87% on ERQA, a test of embodied reasoning, which means understanding space, objects, and how things move. That kind of skill matters for robots and for any tool that has to interpret video or images of real scenes.
The third area is science. Google reports 57% on RareBench and 52.2% on LabBench, both using private evaluations from Gamow Labs. These scores look low next to the 87% above, and the gap makes sense, since tests like these are built to be hard. Again, these are Google's reported numbers and the tests are private, so outsiders cannot easily check them.
The fourth area is coding, especially agentic coding, where the model works through a multi-step programming task on its own. Google says shortcut attempts on coding benchmarks fell to 2%. A shortcut attempt is when a model tries to pass a test by gaming it, for instance by special-casing the expected answer, instead of solving the real problem. Google also reports that agentic code submissions in Antigravity went up by more than 35% and that rollbacks dropped by nearly a third. Antigravity is one of the products that gives access to Argon, and we cover it in the access section below.
Science and robotics examples from Google
Google gave three concrete examples outside of benchmarks, and they are the most readable part of the announcement.
In chemistry, Google says Argon helped with xTB quantum chemistry work and delivered structure prediction up to 2.3 times faster. Structure prediction means working out the shape a molecule is likely to take, which is a slow step in a lot of chemistry research. The phrase up to 2.3 times faster describes a best case, so expect the gain to vary by task.
In biology, Google says the model worked on enzyme design and produced five novel serine esterase designs that were confirmed in the lab. The lab confirmation is the part to notice. Plenty of AI systems can propose a design on screen, and far fewer can point to physical results that held up when scientists tested them.
In robotics, Google points to the Apptronik Apollo humanoid robot sinking seven cornhole shots from a simulation built from video. In plain terms, a simulation of the task was created from video footage, and the robot then used it to land seven shots. It is a demonstration of the model reasoning about physical movement, and one demo does not tell you how a robot will do on other jobs.
None of these examples tell a small business owner what Argon will do for a spreadsheet or a customer email. They do show where Google thinks the model is strongest, which is difficult technical work with checkable outcomes.
Who uses Argon and how to get it
Access depends on who you are and what you pay, and Google splits it into several groups.
Developers can use the Gemini API on the paid tier, through AI Studio and Android Studio. Enterprises can reach Argon through Gemini for business, the API, and Google Antigravity. Consumers get it through subscription plans. Ultra subscribers can use Argon in Antigravity, AI Studio, the Gemini app, and AI Mode in Search. Pro subscribers can use it in the Gemini app and AI Mode in Search.
That means a person on the free plan does not get Argon. The free plan runs on Flash-Lite, a lighter model. If you want to compare the family step by step, our piece on Gemini 3.8 Flash covers the previous generation's fast model and gives useful context for what Flash-class models are good at.
For a small team, the practical routes are the Pro subscription for people who chat and research, and the API for anyone building Argon into a product or an internal tool. Antigravity matters most if someone on your team writes code with AI help.
Consumer plans and prices in the US
Google's subscription page lists these US plans. The page is at gemini.google/us/subscriptions, and prices can change, so check it before you buy.
Free costs $0 and uses Flash-Lite. AI Plus costs $4.99. AI Pro costs $19.99 and includes Gemini 4 Argon access, four times the usage of the Free plan, and 5 TB of storage. AI Ultra starts at $99.99 and gives five times the usage of Pro, and a higher Ultra tier at $199.99 gives 20 times the usage of Pro.
The jump between Pro and Ultra is large. Ultra at $99.99 costs about five times what Pro does and offers five times the Pro usage, so the cost per unit of usage is roughly the same, and what you buy is room to run heavy workloads. Most individuals will find that Pro at $19.99 is the entry point for Argon. Ultra makes sense for people who hit usage limits daily or who need the Antigravity and AI Studio access that comes with it.
AI Plus at $4.99 is the cheaper step above Free, but the information on Google's page that we are working from does not list Argon access for it. If Argon is the reason you are paying, plan on Pro unless Google's page says otherwise when you read it.
What the API costs
For developers, Google prices Argon per million tokens. A token is a small chunk of text, about three quarters of a word in English. Input tokens are what you send, including instructions and documents. Output tokens are what the model writes back.
The introductory price is $2 per million input tokens and $10 per million output tokens. Cached input, which is text the model has already processed and you send again, is 95% off the input rate. By our arithmetic that works out to $0.10 per million cached input tokens at the introductory price. From November 30, 2026, the price rises to $4 per million input tokens and $20 per million output tokens, double the introductory rate.
The date matters if you are budgeting. A project that looks cheap in October may cost twice as much in December, so build your estimates on the higher numbers unless you can finish before the change. If you want to see how these prices stack up against other providers, our API pricing comparison for October 2026 lines them up. For a rival launch from the same month, see our explainer on OpenAI's GPT-6.1 Sol, which is priced at $2 input and $10 output per million tokens.
A worked cost example
Suppose your team runs a document review job that, across many separate requests, sends 1 million input tokens and receives 200,000 output tokens. At the introductory price, input costs $2.00 and output costs 0.2 million times $10, which is $2.00. The total is $4.00.
After November 30, the same job costs $4.00 for input and 0.2 million times $20, or $4.00, for output. The total is $8.00. If you run that job 500 times a month, the bill goes from $2,000 to $4,000 at the new rates.
Caching changes this when your requests share a long common start, such as the same policy document or the same instructions each time. If half of the input were cached at the introductory rate, that half would cost $0.10 per million instead of $2, and the input line would drop from $2.00 to about $1.05. These figures are our calculations from Google's published prices, so treat them as estimates and check your actual usage in the billing dashboard.
One more detail affects planning. Google lists a 1 million token limit through the Gemini API, and it is a unified input and output limit. That means input and output share the same budget. If you fill the whole million with input, there is no room left for the model to answer. Leave space for the response when you send very large documents.
Safety measures Google describes
Google spends a good part of its announcement on safety, and the claims are specific enough to list. The model was evaluated under Google's Frontier Safety Framework. In external red teaming for biological risks, where outside experts try to push the model into giving dangerous help, Google says the model unlocked under 1% of sub-tasks. Google also says harmful-request violations were nearly halved compared with earlier Gemini models.
On prompt injection, which is an attack where hidden instructions inside a web page or document try to hijack an AI agent, Google reports a 0.7% attack success rate in a test by Gray Swan using 15 adaptive attempts. Adaptive means the attacker changes tactics after each failure. This matters if you plan to let a model browse or read files on your behalf, because those are the situations where injected instructions show up.
Google also says the model is trained to pause for human confirmation before irreversible steps, and that an automated misalignment monitor watches its behavior. The first of these is useful for ordinary users. If you set up an agent that can send emails or delete files, a model that stops to ask before doing something permanent gives you a chance to catch a mistake.
Two access programs round out the section. The Fairwind Program lets vetted cyber defenders apply for access without cyber guardrails. Separately, 10,000 scientists can apply for six months of free top-tier Antigravity.
Caveats before you commit
Most of the performance data in this article comes from Google. The Vals Index ranking, the ERQA score, the Gamow Labs results, and the prompt injection figure were all reported by Google in its own announcement. Independent testing may land in different places, and we have not run these tests ourselves.
The scores also describe specific test setups. A 2% shortcut rate on coding benchmarks does not guarantee that the model never cheats on your codebase, and a 0.7% injection success rate across 15 attempts does not mean an agent is safe to leave unattended with sensitive accounts. Keep a human in the loop for anything that touches money, customer data, or permanent changes.
Price is the second caveat. The API rate doubles on November 30, and consumer plan prices and limits are set by Google and can change. The third is access. Argon sits in Pro and Ultra, not in Free, so a person trying the Gemini app on a free account will not see it.
Finally, the Gemini 4 family is just starting, so other models may follow. If you are not in a hurry, waiting to see the rest of the family could give you more options. If you need the top model for a project now, the introductory pricing window is the cheapest time to start.
How to try Argon without overspending
Start with a test set of real work. Pick 20 to 30 tasks from your own files, such as emails to summarize, contracts to review, or spreadsheets to explain, and run them through Argon. Mark each answer as acceptable or not. That gives you a pass rate that means more than any benchmark chart.
If you are an individual, try Pro for a month and cancel if you do not use it. If you are a developer, set a spending cap in the API console, use caching for repeated instructions, and keep your prompts short. Put the stable parts of a prompt at the start and the changing parts at the end so more of each request qualifies for the cache discount.
Teams choosing between providers should run the same test set on each one. Our ChatGPT, Claude, and Gemini comparison is a good place to begin framing what to test, since the best choice often depends on the kind of work more than on the model's headline ranking.
Frequently asked questions
When was Gemini 4 Argon released?
Google announced Gemini 4 Argon as its latest frontier model, globally available from October 13, 2026. Google describes it as just the beginning of the Gemini 4 family.
How much does the Gemini 4 Argon API cost?
The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. From November 30, 2026, the price rises to $4 per million input tokens and $20 per million output tokens.
Which Google plans include Gemini 4 Argon?
AI Pro at $19.99 includes Gemini 4 Argon access in the Gemini app and AI Mode in Search. AI Ultra, from $99.99, adds Antigravity and AI Studio. The Free plan uses Flash-Lite and does not include Argon.
What is the context limit?
Google lists a 1 million token unified input and output limit through the Gemini API. Input and output share that combined limit, so a very large prompt leaves less room for the answer.
Are Google's benchmark results independent?
No. The Vals Index ranking, the ERQA score, the Gamow Labs results, and the Gray Swan prompt injection figure are results reported by Google. Independent testing may differ, so run your own examples before relying on them.
Can I get Argon for free?
Not through the consumer Free plan, which runs on Flash-Lite. Google does say 10,000 scientists can apply for six months of free top-tier Antigravity, and vetted cyber defenders can apply to the Fairwind Program.