On 30 September 2026, Google announced Gemini 4 Argon, its new frontier AI model. On paper it is the strongest model Google has ever shipped: it leads most of the benchmarks Google published, it can write up to one million tokens in a single answer, and it costs a fraction of OpenAI’s flagship.
There is a catch. You cannot use it today. Neither can most developers. This article explains what Argon is, what is new, how it compares with GPT-6 Astra and Claude Opus 5.5, what it will cost in rupees, and when ordinary users are likely to get it.
What is Gemini 4 Argon?
Argon is the first model in Google’s Gemini 4 generation. Google describes it as built for “deep reasoning across complex, long-horizon workflows”. In plain language: tasks that take hours of steady work rather than one quick answer, such as migrating a large codebase, auditing software for security holes, or researching and writing a long report.
It is Google’s first flagship model in months. According to 9to5Google, it follows the cancellation of Gemini 3.5 Pro, with Google focusing on the lighter Gemini 3.8 Flash in between. During that gap OpenAI released GPT-6 Astra and Anthropic released its Claude 5 models, so Argon is Google’s answer to both.
The name is new as well. Instead of “Pro” or “Ultra”, the model carries the name of an element. Google has not explained the naming or said which other Gemini 4 models are coming.
What is new
1. One million output tokens
This is the biggest technical change. Earlier Gemini models could write about 64,000 tokens in one response. Argon can write up to 1 million.
A token is a small piece of a word; 1 million tokens is very roughly 7 lakh English words. Why does this matter? Because the output limit also caps how long a model can “think” and how much it can produce in one go. With 64,000 tokens, a model rewriting a large program has to stop and be restarted many times. With a million, it can finish the whole job in one run.
Google gave an example: Argon replaced 32,000 lines of hand-tuned code in an open-source video decoder (libgav1), and the result ran 2.7 times faster than an existing Rust version.
2. Built for cybersecurity defence
Google says Argon can autonomously find, validate and patch critical software vulnerabilities. In early use through security company Wiz’s “Scan for Good” initiative, it found a critical flaw in a healthcare system that exposed sensitive personal information.
3. Better at long videos and documents
On LVBench, a test of understanding long videos, Argon scored 91.7%, the highest reported so far. It can also read charts and other visual material.
4. Harder to trick
AI agents that browse the web or read emails can be hijacked by hidden instructions planted in that content. This is called indirect prompt injection. On Gray Swan’s test for this, attacks succeeded against Argon only 0.7% of the time, compared with 1.0% for Claude Opus 5.5 and 8.5% for GPT-6 Astra.
Benchmarks: Argon vs GPT-6 Astra vs Claude Opus 5.5
These are the numbers Google published, as reported by VentureBeat. Remember that every company picks the tests it shows, so treat this as Google’s side of the story until independent results arrive.
| Benchmark (what it tests) | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE v1.1 (software engineering) | 77.9% | 74.1% | 74.2% |
| AutomationBench (business workflows) | 51.3% | 41.4% | 42.5% |
| GraphWalks (long-context reasoning) | 84.2% | 71.8% | 66.8% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% |
| Harvey Legal Agent | 19.6% | 5.4% | 3.8% |
| LVBench (long video) | 91.7% | 87.5% | 83.7% |
| CWE-bench v1 (finding vulnerabilities) | 68% | 68% | 67% |
| FrontierSWE v2 | 55.0% | 65.5% | — |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | — |
| Terminal-Bench 4.0 | 57.4% | — | 66.4% |
| PostTrainBench | 45.3% | — | 49.3% |
In total, Argon leads or ties in 13 of the 18 categories Google disclosed. It does not win everything: GPT-6 Astra is ahead on two hard engineering and science tests, and Claude Opus 5.5 is ahead on terminal-based coding. Benchmarking startup Vals also places Argon at the top of its own model index.
Notice the Harvey Legal Agent row. Even the winner scores under 20%. These models are improving quickly, but long, expert-level professional work is still far from solved.
Price: the real surprise
| Model | Input (per 10 lakh tokens) | Output (per 10 lakh tokens) |
|---|---|---|
| Gemini 4 Argon (introductory) | $2 (about ₹188) | $10 (about ₹940) |
| Gemini 4 Argon (after introductory period) | $4 (about ₹376) | $20 (about ₹1,880) |
| Claude Opus 5.5 | $4 (about ₹376) | $20 (about ₹1,880) |
| GPT-6 Astra | $10 (about ₹940) | $50 (about ₹4,700) |
At the introductory rate, Argon costs one-fifth of GPT-6 Astra. Repeated (cached) input gets a 95% discount. For an Indian startup running an AI coding assistant or a customer support agent, model cost is often the biggest monthly bill after salaries, so a five-times difference decides which model gets used.
Two cautions. The introductory price will double later, and Google has not said when. And a model that can write a million tokens can also run up a large bill in a single request, so output limits in your own code will matter.
So why can nobody use it?
For now Argon is going only to trusted cybersecurity defenders through Google’s Fairwind Program, and to Google’s own teams. Those partners get the model without the usual cyber guardrails, so that they can use its full ability to find and fix vulnerabilities. Google is also taking part in the US government’s voluntary pre-release model testing process.
The reasoning is simple. A model that can find and patch security holes can also find and exploit them. Giving defenders a head start before the public gets access is meant to let them fix the worst problems first. OpenAI did something similar with GPT-6 Astra, which reached the public only in a version that refuses certain cybersecurity requests.
Google lists other safeguards too: monitoring the model’s internal activations for signs of misuse, running agents inside hardened sandboxes, and watching the model’s reasoning with the ability to stop a task midway.
When will it reach India?
Google says wider access will come “as soon as possible”, starting with paid API customers and Google AI Ultra subscribers. No date has been given.
- Developers: expect it in the Gemini API for paid accounts first. Free-tier API keys are unlikely to get it early.
- Gemini app users: Google AI Ultra, which is reported to start at ₹6,500 a month in India, is first in line. There is no word yet on the cheaper plans or on free users.
- Students and free users: you will most likely keep using Gemini 3.8 Flash for now.
What should you do now?
- Do not switch tools yet. You cannot test Argon, and Google’s benchmark table is not the same as your own experience.
- If you build with AI APIs, keep your code model-agnostic so that you can try Argon the day it opens. At ₹188 per 10 lakh input tokens it will be worth a test.
- If you work in security, watch the Fairwind Program. Models like this are about to change how vulnerabilities are found, on both sides.
- Wait for independent tests. The interesting question is whether Argon’s lead holds up on tasks Google did not choose.
The bottom line
Gemini 4 Argon puts Google back at the front of the AI race on most published measures, at a price that undercuts its rivals. But a model nobody can use is a promise, not a product. The real test comes when Indian developers and Gemini users can try it on their own work. Until then, the fair summary is: very strong claims, an aggressive price, and a release date that is still blank.
Comments (0)
Be the first to comment!