From coding to cyber defence: What makes Google's Gemini 4 Argon different

Google says the longer context helps Gemini 4 Argon handle complex, multi-step tasks by retaining information and decisions across a single workflow instead of splitting the work into separate session

Google has introduced Gemini 4 Argon, its frontier artificial intelligence model focused on handling complex, multi-step tasks. According to Google, the model can work across software engineering, financial and legal research and cybersecurity, with the ability to sustain reasoning over much longer tasks. A key change is its output limit of up to one-million tokens, compared with 64,000 tokens previously.

 

Google is initially giving access to trusted cyber defenders while it continues testing the model’s safety systems before a broader release. The cautious rollout reflects a wider push by AI companies to add safeguards and restrictions before releasing more capable models.

 

 

What is Gemini 4 Argon

 

Gemini 4 Argon is designed for what Google describes as “long-horizon” work. It means that the model is intended to continue working through a complicated task involving multiple steps instead of stopping after producing a short answer.

 

This is relevant for tasks such as writing and debugging software, analysing financial information, conducting legal research or finding security vulnerabilities in software. These tasks can require a model to examine a large amount of information, make decisions at different stages and continue working until it reaches an outcome.

 

When will Gemini 4 Argon be available

 

Gemini 4 Argon is not being released as a general-purpose model immediately. Google says it is first rolling out the model to a group of trusted cyber defenders and testers. Their feedback will be used to improve the safeguards before the company expands access.

 

The eventual rollout will begin with paid API customers and Google AI Ultra subscribers, according to the company.

 

Google has also announced introductory API pricing of $2 per million input tokens and $10 per million output tokens. After the introductory period, the prices will increase to $4 per million input tokens and $20 per million output tokens.

 

Google's decision to begin with trusted cyber defenders reflects that tension. The company is expanding access while continuing to test safeguards before making Argon available more widely.

 

Why the 1 million-token limit matters

 

One of the biggest changes in Argon is the amount of output it can generate in a single trajectory. Google has increased the model's output token limit to 1 million tokens from 64,000 tokens. A token is a basic unit of text processed or generated by an AI model. The exact number of words represented by a token varies, but a 1-million-token limit gives the model considerably more room to work through a large problem.

 

For long-running tasks, this matters because the model does not have to stop as frequently or divide its work into as many separate sessions.

 

Google says the additional space allows Argon to reason more deeply and generate hundreds of thousands of tokens during a single task. The longer limit helps when a task involves multiple steps, as the model can retain earlier information and decisions while working through the problem.

 

The change is therefore not simply about producing longer answers. It is about allowing the model to maintain a longer chain of work while solving a problem.

 

For context, Anthropic’s Claude Opus 5.5 supports up to 300,000 output tokens and OpenAI’s GPT supports up to 128,000 maximum tokens per request. 

 

From coding to enterprise work

 

Coding is one of the main areas where Google says Argon is being used. Google engineers are using the model for tasks ranging from everyday debugging to large code migrations and algorithm design. The company says Argon achieved a score of 77.9 per cent on DeepSWE v1.1, a benchmark focused on real-world, long-horizon software engineering tasks.

 

The model is also designed to work beyond software development.

 

According to Google’s blog, Argon leads the Vals Index, which tests work across finance, coding, legal and tax. It also performs strongly in financial research and legal research and drafting.

 

Another benchmark cited by Google is AutomationBench, which measures end-to-end execution across business functions. Argon recorded a score of 51.3 per cent on this evaluation.

 

The model's capabilities are not limited to text. Google noted that Argon can analyse professional charts, identify information from long videos and act on information spread across multiple documents. On LVBench, which measures long-video understanding, Google reports a score of 91.7 per cent.

 

These examples show how Google is positioning Argon as an AI model that can handle complete professional tasks, rather than simply respond to individual questions.

 

Where cybersecurity becomes different

 

The most significant part of the announcement is arguably Argon's cybersecurity capability.

Google mentions the model has been trained to assist cyber defenders and can autonomously find, validate and patch critical software vulnerabilities.

 

Finding a vulnerability is only one part of security work. A security team also needs to establish whether a suspected weakness can actually be exploited, understand its impact and then determine how to fix it. Google says Argon is capable of working through these stages.

 

The company is initially making the model available to trusted cyber defenders through its Fairwind programme. Google says its own internal teams will also receive access to the model without cyber guardrails so they can use its full cybersecurity capabilities for defensive work.

 

One early example comes from Wiz's Scan for Good initiative, which uses Argon to identify and remediate high-risk exposures affecting critical public infrastructure.

 

Google says the model uncovered a critical vulnerability affecting healthcare software used by hospitals around the world. According to the company, the vulnerability exposed sensitive personal information and had not been identified by earlier frontier models.

 

Google also reports that Argon tied for first place on CWE-bench v1, a benchmark for security vulnerability remediation, with a score of 68 per cent. 

 

Monitoring what the model does

 

Google is also focusing on what it calls misalignments, situations in which a model could take actions that go beyond what the user intended while trying to complete a task.

 

For Argon, Google says it is deploying systems that monitor the model's reasoning and actions and can stop execution when necessary. The company says it used a similar monitoring approach during training. The monitoring system could alert a dedicated incident response team when it detected potential problems.

 

Google is also hardening the environments in which powerful models are tested. Its approach includes isolating and securing sandboxed environments before high-risk training or evaluations begin.

 

This is relevant as AI systems become more capable of taking actions rather than simply generating text. A model that can write code, interact with systems and work through a task autonomously requires more controls around what it can access and what happens if its behaviour goes off course.