Google unveils Gemini 4 Argon and reignites the fight for the AI frontier

On September 30, Google did something that had been pending for months: introduce another model designed to compete at the highest end of the market.

It's called Gemini 4 Argon and, curiously, Gemini 3.5 Pro fell by the wayside. Google confirmed that the model it had previously announced will no longer be released. Argon becomes the family's new benchmark directly.

Beyond the version jump, I found several things that are actually worth looking at. One is an output limit of 1 million tokens. Another is that Google is already using Argon within its own infrastructure for work that goes well beyond generating text.

And there's a third: we still can't use it freely.

One million tokens changes the type of problem

In the previous generation, the output limit was 64 thousand tokens. With Argon it goes up to 1 million.

That represents an increase of more than 15 times.

It doesn't mean it now makes sense to ask for a million tokens every time we run a query. That would be absurd. What's interesting is giving the model room when it works as an agent during long processes.

Google talks about exactly that: tasks with many steps, deep research, software engineering and business processes that may require analyzing documents, executing actions and continuing to reason for quite a while.

It's similar to giving a developer a much bigger notebook. It doesn't guarantee they'll write better solutions, but they stop running out of pages halfway through the problem.

Google officially describes Argon as a model designed for deep reasoning in long and complex workflows.

Google already put it to work at home

This was probably the part of the announcement that interested me most.

Google states that thousands of its employees already use Argon for internal tasks related to programming, research and writing.

One of the examples directly involves its data centers.

Argon-based agents analyzed performance information from Google's infrastructure, identified optimization opportunities and proposed changes that, once implemented, freed up more than 300 TiB of memory. Google estimates the full potential could be between 500 TiB and 1 PiB.

That already seems like a much more interesting case than asking an AI to build an app in a ten-minute demo.

It's also modifying enormous codebases

Google says it's using Argon agents to migrate software written in C and C++ to Rust.

The projects mentioned range from libraries of tens of thousands of lines to components of the Zircon kernel in Fuchsia that exceed 800 thousand lines of code.

Obviously, that doesn't mean Google unleashed an AI on the repository and sent everything straight to production. The company itself explains that these migrations go through automated audits, human review, and testing before being deployed.

That detail matters.

To me, the interesting future of these tools lies precisely there: systems where the agent can do a lot of work, but there's a validation chain around it.

Does it really beat OpenAI and Anthropic?

It depends on what we measure.

And that's one of the parts I liked about the announcement, because the table published by Google includes results where Argon wins and others where it loses.

The results are the ones published by Google DeepMind and should be interpreted as benchmarks, not as a guarantee that Argon will be better on any real project.

On DeepSWE v1.1, focused on long-duration software engineering, Argon scores 77.9%.

But on FrontierSWE v2 it falls quite a bit behind GPT-6 Astra. And on Terminal-bench 4.0, Claude Opus 5.5 has a clear advantage.

Reuters also highlighted precisely this: although Argon beats its competitors on several tests published by Google, it remains below on other metrics, including programming-related tests.

That's how any model launch in 2026 should be read.

There's no single table that tells us which model works best for everything.

Argon's big bet is on cybersecurity

Google devoted a considerable part of Argon's training to cyber defense.

According to the company, it can autonomously locate vulnerabilities, verify whether they're actually exploitable, and generate a fix.

On CWE-bench v1 it scored 68%, tied with GPT-6 Astra.

Even more interesting: Google says Wiz is already using Argon within its Scan for Good initiative. During testing, the system identified a critical vulnerability that exposed personal information in software used by hospitals and that other frontier models had not found.

That capability also explains why Google hasn't fully opened up the model yet.

A system that's very good at finding vulnerabilities can serve both to defend infrastructure and to look for ways to attack it.

For now, limited access

Gemini 4 Argon is beginning its rollout through the Fairwind program, initially aimed at cybersecurity specialists considered trusted.

Google is using this period to test protections against abuse, prompt injection attacks, and agent behaviors that might stray from the original instructions.

The company says access will later begin for paying customers of its API and Google AI Ultra subscribers. It did not provide an exact date for the general public launch.

It also announced pricing.

During the introductory period:

• Input: $2 per million tokens
• Output: $10 per million tokens
• Cached input: 95% discount

After the promotional period, Google says the price will rise to $4 input and $20 output per million tokens.

It's a pretty clear sign that Argon is meant first for complex work where the model's cost is justified by the value of the task.

Google needed this launch

Gemini had been in a strange position for months.

Google kept releasing Flash models and improving its ecosystem, but its next big model was taking a while as OpenAI and Anthropic continued unveiling frontier models.

Gemini 3.5 Pro was eventually canceled. Reuters confirmed with Google that there are no longer plans to release it.

Argon seems to be the attempt to close that chapter and come back directly with a more ambitious architecture.

The name also marks an interesting change. Instead of sticking solely with Pro, Flash, and similar variants, Google is starting Gemini 4 using Argon to identify its frontier model.

We'll have to see whether this naming continues with other models in the family.

My takeaways

Of the whole announcement, four things stand out to me.

The million output tokens is important, but mainly for agents and long jobs. For a normal conversation it practically makes no difference.

Google's internal cases are more interesting than many benchmarks. Freeing up hundreds of TiB of memory or working on migrations of hundreds of thousands of lines brings AI closer to problems that actually cost money.

Argon doesn't dominate every test. It has excellent results in some and clearly loses in others. That makes it even more necessary to evaluate models with our own work.

And cybersecurity is becoming one of the most delicate frontiers of advanced models. The same tool capable of finding a critical vulnerability to fix it can turn into a problem if it ends up in the wrong hands.

Conclusion

Gemini 4 Argon does seem like an important launch for Google.

Not because a chart says it scored three points higher than another model, but because examples are starting to appear where AI agents work for long periods on real infrastructure, software, and business processes.

Even so, the most important test is still missing.

We developers still don't have general access to Argon.

When that happens, I want to test it with large repositories, scattered documentation, tasks that span multiple stages, and errors that aren't specially prepared for a demo.

That's where we'll see how much is benchmark and how much is an improvement that truly changes the way we work.

Main sources: Google: Gemini 4 Argon · Google DeepMind: Gemini 4 Argon · Reuters: Gemini 4 Argon announcement