Gemini 4 Argon: Benchmarks, Pricing, and Why Cyber Defenders Got It First

Google announced Gemini 4 Argon on September 30, 2026, and framed it as the company's next era of frontier intelligence. The headline specs are real: a context window that moved from 64,000 tokens to one million, a number one rank on AutomationBench at 51.3 percent, and a tied for first result of 68 percent on CWE bench v1, a benchmark that measures whether a model can find and patch real vulnerabilities rather than just discuss them. What makes this release worth a closer look is not only the scores. It is who gets to use the model first.

Rather than a simultaneous public launch, Google is rolling Argon out through its Fairwind Program, which prioritizes trusted cybersecurity defenders ahead of the broader developer and consumer audience. A U.S. government voluntary pre release process is running in parallel. That sequencing is a deliberate signal about what Google thinks this model is actually for, and it has direct implications for anyone evaluating frontier models for enterprise use this quarter.

The benchmark numbers

Benchmark	Score	What it measures
DeepSWE v1.1	77.9%	Real world software engineering task completion
AutomationBench	51.3% (rank 1)	Multi step workflow automation
CWE bench v1	68% (tied for first)	Vulnerability detection and remediation
LVBench	91.7%	Long video understanding
Vals Index	Leading across finance, coding, legal, tax	Composite professional task performance
Gray Swan IPI	Leading	Robustness against prompt injection attacks

Google's own framing claims Argon scores significantly higher than OpenAI's GPT 6 Astra and Anthropic's Fable and Opus model families across its benchmark suite, and cites the independent evaluator Vals as ranking Argon first on its model index. Treat vendor comparisons like this the way you would treat any single sourced claim: useful as a signal that the company is confident in a head to head, not as a substitute for testing the model against your own workload. Vals is a third party, which gives the claim more weight than a purely internal number, but it is still one evaluator's methodology.

Why the context window jump matters more than the headline score

A 64,000 token context window to a one million token context window is a sixteen fold increase, and it changes what kinds of tasks are even possible without a retrieval layer in front of the model. At 64,000 tokens, a model can hold a few source files and a conversation history. At one million tokens, it can hold a meaningful slice of an actual codebase, a long incident timeline, or an extended cybersecurity investigation log, all in a single context, without chunking and without the retrieval infrastructure that chunking requires. For the vulnerability remediation use case Google is explicitly targeting with the Fairwind rollout, that matters a great deal: a security review of a real system usually involves cross referencing logs, configuration files, and source code that together exceed what a 64,000 token window can hold at once.

Pricing: what it actually costs to run

Google set introductory pricing at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached input tokens. After the introductory period, standard pricing moves to $4 per million input and $20 per million output. The cached input discount is the detail worth paying attention to if you are running Argon against a large, stable context, such as a fixed codebase or a standing set of security policies: a workflow that repeatedly queries the same million token context pays the full rate once and the heavily discounted rate on every subsequent call that reuses the cached portion. That economics only works if your actual usage pattern is high context reuse rather than constantly shifting context, so model this against your own call pattern before assuming the sticker price applies to your workload.

Why cyber defenders got it first

Google's stated rationale centers on letting cybersecurity professionals safely validate the model's vulnerability detection and patching capabilities in a controlled setting before wider release. The safety framework paired with the release includes misuse prevention, prompt injection defense, misalignment monitoring, and hardened sandboxed testing environments, which reads as Google acknowledging that a model capable of autonomously finding and patching vulnerabilities is also, by construction, a model capable of finding and exploiting them if it ends up in the wrong hands or is misused by a legitimate one.

This is a meaningfully different release pattern from a standard frontier model launch, and it arrives in the same week that OpenAI pulled a planned model release over its own safety findings. Two major labs making safety gated decisions about high capability models in the same seven day window is not a coincidence worth ignoring if you are building a vendor evaluation process: it suggests staged, capability gated releases are becoming the norm for models with autonomous security or coding capability, not the exception.

What this means if you are choosing a model for enterprise work

If your evaluation criteria include raw benchmark performance, Argon is a legitimate contender on the numbers Google has published, particularly for long context, code heavy, or security adjacent workloads. If your evaluation also includes availability and support maturity, you are not evaluating the same model as the cyber defenders currently testing it. Broader access is coming, per Google's own framing, but "coming soon" is not a date you can put in a procurement timeline. The practical move for most teams is to treat Argon as a model to watch and benchmark against your own tasks the moment general availability opens, while continuing to build on models you can access today, rather than waiting on a staged rollout with no published end date.

When this applies to your stack

If you are running agentic workflows that depend on long context, cross referencing large codebases, extended logs, or standing policy documents, Argon's context window and cached pricing are worth modeling against your actual token usage once general availability opens. If your team is evaluating multiple frontier models for a coding agent or security tooling build, the AI integration and agent engineering work of wiring a new model into an existing pipeline without a rewrite is exactly the kind of migration Contra Collective handles for clients moving between model providers as the frontier shifts.

FAQ

Is Gemini 4 Argon publicly available yet? No. As of this writing, Argon is rolling out through Google's Fairwind Program to trusted cybersecurity partners and through a voluntary U.S. government pre release process. Google has said broader availability is coming but has not published a general availability date.

How does Gemini 4 Argon's pricing compare to the previous Gemini generation? Google has not published a direct prior generation comparison alongside this release. The introductory rate of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period, is the number to benchmark against whatever model you are currently running once you can access Argon directly.

What does the CWE bench v1 score actually measure? CWE bench v1 tests whether a model can identify a real Common Weakness Enumeration vulnerability in code and produce a working patch for it, not just describe the vulnerability class. Argon's 68 percent, tied for first, means it successfully remediated roughly two out of every three tested vulnerabilities.

Why does the one million token context window matter for security work? A security investigation often spans logs, configuration, and source code that together exceed a small context window. A one million token window lets a model reason over a much larger slice of that material in a single pass, without the retrieval chunking that smaller windows require and that can drop relevant detail between chunks.