Google Releases Gemini 4 Argon, Topping Multiple Leaderboards and Focusing on Cybersecurity Defense

QbitAI | WeChat Official Account QbitAI

Oh wow, Google has finally gotten its act together...

On the first day of the holiday, Google dropped Gemini 4 Argon and it shot straight to the top of multiple leaderboards.

It punches out Opus 5.5 and kicks GPT-6 Astra.

And the wilder part is still to come: the cost per individual task can be as low as $1.99, directly halving Astra's cost.

A maximum output limit of one million tokens, aimed at complex workflows in programming, finance, and law, and — mark this — its cybersecurity defense capabilities are absolutely stellar.

The wait-and-see crowd is about to win big.

But, well, ordinary users like us can only admire it from afar for now; we can't get our hands on it just yet.

Gemini 4 Argon is currently only open to a select group of vetted cybersecurity teams; other subscribers will have to take it step by step.

That said, Google's internal attitude toward the new-generation flagship model is rather interesting.

Sundar Pichai, Hassabis, and successor Koray Kavukcuoglu all gave it a one-click triple (like, repost, comment) in support: Google is back!

DeepMind researcher Zirui Wang posted bluntly:

RSI is here! There's nothing left for me to do; it's time to chase my barista dream :)

For real?

Everyone stand up! The Gemini 4 we've been waiting for with bated breath has finally arrived.

Count carefully — it's been almost a year since Google's previous flagship model launched; even Nezha wasn't this difficult to deliver.

Over this past year, OpenAI and Anthropic have been chasing and slashing at Google; the week before the holiday they even released Opus 5.5, Sonnet 5.5, and GPT-6.1 Sol in succession to lay siege.

You think Google isn't anxious? The new model Argon was rushed ashore at top speed.

First, the benchmark results: a clean sweep of performance leadership.

Argon scored 77.9% on DeepSWE v1.1, which measures long-horizon software engineering tasks, above Claude Opus 5.5's 74.2% and GPT-6 Astra's 74.1%.

On the cybersecurity vulnerability remediation test CWE-bench v1, it scored 68%, tying for first with GPT-6 Astra.

Argon topped the Vals Index at 68.90%, a substantial improvement over Gemini 3.8 Flash.

In particular, its RSI index jumped to fourth, surpassing GPT-6 Astra.

Gemini 4 Argon is no slouch in other third-party evaluations either.

In the text domain, Argon shows significant advantages in code writing, high-difficulty prompts, instruction following, long queries, and creative writing, ranking first in Text Arena with 1525 points.

In Code Arena, which evaluates web development, it ranks eighth, up 21 places from Gemini 3.8 Flash High, but still slightly behind the latest Claude and GPT models.

In the Artificial Analysis evaluation, Argon is on par with GPT-6 Astra, one point above GPT-6.1 Sol, with the improvement mainly coming from fewer hallucinations and stronger agentic capabilities.

But on price, Argon has a clear advantage over Astra.

During the promotional period, Argon's cost per question is about 40% lower, at $2 per million input tokens and $10 per million output tokens, making it currently one of the most cost-effective models.

In concrete work, Argon also leans toward handling complex, long-horizon deep reasoning tasks, such as software engineering, enterprise expertise organization, and cybersecurity defense.

The model's one-million-token output limit is precisely for such long tasks, giving the model more room to work continuously and ample reasoning depth.

Take Google's internal usage as an example: Argon autonomously identified and implemented memory optimization plans across Google's data centers, and once fully deployed, the plans are expected to free up over 300 TiB of memory.

In another case, the Argon Agent, working on a Rust port of a video decoder, repeatedly experimented and replaced 32,000 lines of SIMD code, with the final version running at 2.7 times the speed of the original Rust port.

Officials also emphasized Argon's defensive cybersecurity advantages.

It can autonomously discover, verify, and patch critical software vulnerabilities in more than 20 languages; in Wiz's internal black-box penetration testing, the model could analyze network systems in real time without source code, efficiently discovering attack surfaces, identifying vulnerabilities, and generating proof-of-concept evidence.

To prevent abuse, Argon first enters Google's Fairwind program, aimed at vetted cybersecurity defenders.

For trusted defense teams, Google will lift some restrictions on cybersecurity tasks, and Argon will help defenders patch holes and find attack entry points.

But talk is cheap; how Argon actually performs in practice remains a question mark.

After all, the general public hasn't gotten to use it yet...

There are rumors that Google employees are not entirely optimistic about Gemini 4.

Some employees who have had access to Gemini 4 believe that the model's real-world programming performance is not as impressive as its benchmark scores, that front-end interface design for web pages and apps remains a weak spot, and that there is even suspicion of excessive benchmark gaming.

However, another employee familiar with the development work said that internally it is widely believed that Gemini 4 has reached the frontier level. Google also responded that saying its programming performance is poor is inaccurate.

In short, how the model really is remains to be seen.

Fortunately, Google has finally returned to the table.