Google has announced Gemini 4 Argon, its newest frontier AI model, and it is not rolling out to the general public just yet. Instead, Google is starting with a narrow group of “trusted cyber defenders” through what it calls the Fairwind Program, before gradually widening access to developers, enterprises, and everyday users. The move matters because Argon is being positioned not just as a smarter chatbot, but as a model capable of sustaining deep reasoning over long, complex workflows in software engineering, legal and financial analysis, and cybersecurity defense.

What Gemini 4 Argon actually is

Argon is Google’s latest large language model, built to handle long-horizon tasks that require many steps of reasoning rather than quick, one-shot answers. According to Google, the model is already being used internally, with thousands of Google employees relying on it for specialized coding tasks, deeper research, and writing. The company says Argon is “fundamentally changing the way we work and build at Google,” which signals that this release is meant to be a significant step up from previous Gemini models rather than an incremental update.

One of the most concrete technical changes is the model’s output capacity. Google is expanding Argon’s output token limit to 1 million tokens, up from 64,000 tokens in earlier models. In practical terms, tokens are the chunks of text a model reads and generates; a much higher output limit means Argon can produce far longer, more detailed responses without being cut off mid-task. Google frames this as giving the model “headroom to think deeply,” allowing it to work through difficult problems in a single continuous attempt instead of breaking them into multiple exchanges.

Why benchmark performance matters here

Google backs up its claims with results from several independent and third-party benchmarks. On DeepSWE v1.1, which tests how well a model handles realistic, long-horizon software engineering tasks, Argon reportedly sets a new state of the art with a score of 77.9%. On the Vals Index, which measures economic impact across finance, coding, legal, and tax work weighted by contribution to U.S. GDP, Argon is described as the leading model. It also reportedly performs strongly on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark, both of which test multi-step professional research and drafting rather than simple question answering.

Argon also ranks first on Zapier’s AutomationBench, a benchmark for end-to-end execution across core business functions, with a score of 51.3%. Beyond text-based tasks, Google highlights the model’s visual understanding, pointing to a score of 91.7% on LVBench, a test of long video comprehension. Together, these numbers suggest Google is trying to prove Argon’s value not through generic chat quality, but through its ability to complete real, multi-step professional work: debugging codebases, researching legal questions, analyzing financial documents, and interpreting charts or video content.

A cybersecurity-first rollout strategy

What sets this launch apart from a typical model release is the initial focus on cybersecurity defense. Google says it specifically trained Argon to be highly capable at finding, validating, and patching critical software vulnerabilities on its own. For the trusted defenders in the Fairwind Program, as well as Google’s internal teams, Argon will reportedly be released without cyber guardrails, meaning its full offensive-adjacent security capabilities are available for defensive use rather than being restricted the way they might be for general consumer access.

Security firm Wiz is already using Argon through its Scan for Good initiative, a program aimed at protecting critical public infrastructure for free. This gives an early real-world example of how the model’s vulnerability-finding abilities are being applied before wider release. Google also notes it is participating in the U.S. government’s voluntary process for pre-release model access, a step that reflects growing scrutiny around powerful AI systems that can autonomously probe and patch software, since such capabilities could theoretically be misused if they fell into the wrong hands before adequate safeguards are in place.

Pricing and what it means for developers

Google has already published introductory pricing for Argon, even though broad access is not yet available. The model will launch at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at a 95% discount. For developers who already work with large language model APIs, this pricing structure will look familiar, and the steep discount on cached inputs suggests Google is encouraging workflows that repeatedly reference the same large documents or codebases, a common pattern in enterprise knowledge work and long software engineering sessions.

For now, there is no public API or consumer app access to try Argon directly. Google says it will keep gathering feedback from early testers in the Fairwind Program and continue refining guardrails before expanding availability “as soon as possible,” but no firm date has been given. Developers and businesses interested in enterprise coding, legal, or finance applications should watch for Google’s follow-up announcements, since the phased approach means general access could still be some time away.

How Argon fits into the broader AI landscape

The emphasis on long-horizon reasoning, a massively expanded output limit, and benchmark-driven claims across coding, finance, legal, and security work reflects a broader industry trend: frontier AI labs are increasingly competing on how well models handle sustained, multi-step professional tasks rather than short conversational exchanges. Argon’s cybersecurity angle in particular stands out, since autonomous vulnerability discovery and patching is a capability most major labs have approached cautiously. Google’s decision to release this feature first to a controlled group of defenders, paired with government engagement, suggests the company is trying to balance the practical benefits of the technology against the risks of releasing such capable security tools broadly and immediately.

Frequently asked questions

Can I use Gemini 4 Argon right now? Not yet for most people. Access is currently limited to trusted cyber defenders through the Fairwind Program, with wider developer, enterprise, and consumer availability planned to follow in phases.

What makes Argon different from earlier Gemini models? Its output token limit has jumped from 64,000 to 1 million tokens, and it has been specifically trained for long, multi-step reasoning in software engineering, enterprise knowledge work, and cybersecurity defense, with strong results on benchmarks like DeepSWE v1.1 and the Vals Index.

Why is cybersecurity the first use case highlighted? Google trained Argon to autonomously find, validate, and patch vulnerabilities, and is releasing it without cyber guardrails to trusted defenders so they can use its full defensive capabilities, while working through a government pre-release access process to manage risk.

Source: Gemini 4 Argon