Gemini 4 Argon Explained: What Google's New AI Model Can Actually Do

Gemini 4 Argon Explained: What Google's New AI Model Can Actually Do

Quick Answer

Gemini 4 Argon is Google's newest frontier AI model, announced September 30, 2026, built to handle long, complex, multi-step tasks in software engineering, enterprise knowledge work, and cybersecurity defense. Google's own published benchmarks show strong results on coding and vulnerability-remediation tests, and the model now supports up to 1 million output tokens, a major jump from the previous 64,000-token limit. Argon is not broadly available yet — it's currently rolling out only to a small group of trusted cybersecurity defenders through Google's Fairwind Program, with no confirmed date for wider access to developers, businesses, or everyday Gemini app users. Everything about its real-world performance outside Google's own testing and a handful of early partners remains, for now, based on Google's word rather than independent, broad verification.

What Is Gemini 4 Argon?

Argon is Google DeepMind's new flagship model, described in the company's own announcement as built "to sustain deep reasoning across complex, long-horizon workflows." It follows the Gemini 3 generation after a long gap between flagship releases. Reuters reported that Google had previously planned a Gemini 3.5 Pro release before shifting its focus to Gemini 4.

Unlike a typical consumer-focused model announcement, Argon's rollout is explicitly staged. Google says it is participating in the U.S. government's voluntary process for pre-release model access and is gathering feedback from early testers before expanding availability to developers, enterprises, and consumers. 

Confirmed: Coding Capabilities

This is the area with the most concrete, specific detail from Google itself. On DeepSWE v1.1, a benchmark measuring performance on real-world, long-horizon software engineering tasks, Google reports Argon achieving a new state-of-the-art score of 77.9%.

Beyond the benchmark, Google shared specific internal examples of Argon already being used at the company. Argon agents have reportedly been used to migrate large C/C++ codebases to Rust, including Google's re2 and libgav1 libraries and portions of the Fuchsia operating system's Zircon kernel, spanning over 800,000 lines of code. In one specific case involving libgav1, Google's open-source video decoding software, Argon agents reportedly replaced 32,000 lines of performance-critical code through repeated profile-guided experiments, producing a result Google says runs 2.7 times faster than the existing Rust port while preserving identical video output. Google has also stated these large-scale rewrites go through rigorous automated and manual auditing before reaching production, rather than being deployed directly from the model's output.

Confirmed: Reasoning and Context Handling

The most concrete technical upgrade here is the expanded output limit: Google says Argon can generate up to 1 million output tokens in a single trajectory, up from the previous 64,000-token limit. This is an output limit, not a statement that Argon's context window is 1 million tokens.  Google frames this as giving the model "headroom to think deeply" through long, multi-step reasoning chains without needing to split a complex task across multiple separate exchanges.

On broader reasoning and knowledge-work benchmarks, Google reports Argon leading the Vals Index, a benchmark measuring economic impact across finance, coding, legal, and tax work weighted by each sector's contribution to U.S. GDP, along with strong results on Vals' Finance Agent benchmark and Harvey's Legal Agent Benchmark for legal research and drafting. On AutomationBench, a business-task benchmark built by Zapier, Google reports Argon ranking first with a score of 51.3%.

Confirmed: Cybersecurity Capabilities

Cybersecurity is clearly the centerpiece of this launch, and it's also where Google chose to grant its first outside access. Google says Argon was specifically trained to be "highly capable at cybersecurity defense" and can "autonomously find, validate, and patch critical software vulnerabilities." On CWE-bench v1, a benchmark for vulnerability remediation, Google reports Argon tying for first place with a score of 68%.

One real-world example Google cited: the security firm Wiz is already using Argon through its "Scan for Good" initiative, a program aimed at finding and fixing security exposures in public infrastructure software for free. Google says the model uncovered a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide — a risk Google says earlier frontier models had missed. Notably, Google says it will release Argon to trusted defenders "without cyber guardrails" so they can use its full cybersecurity capabilities. Google has not publicly detailed exactly which restrictions are removed, so the practical difference between this version and the broader release is not yet fully documented.

Confirmed: Visual and Multimodal Understanding

Google also describes Argon as "uniquely strong" at tasks requiring visual understanding, including professional chart analysis, extracting details from long videos, and taking action based on a series of documents. On LVBench, a benchmark for long-video understanding specifically, Google reports a state-of-the-art score of 91.7%.

What Google Says It's Doing About Safety

Given Argon's stated cybersecurity capabilities, Google devoted real space in its announcement to safeguards, which is worth summarizing honestly rather than skipping over. Google says it's strengthening defenses against misuse for both cyberattacks and CBRN (chemical, biological, radiological, nuclear) risks, including monitoring the model's internal activations to spot misuse, following its own published Frontier Safety Framework. Google also says Argon is more resilient to indirect prompt injection attacks and describes systems designed to monitor its reasoning and actions for signs of misalignment, with execution stopped when necessary. These are Google's reported safety results and deployment measures, rather than independently reproduced findings.

These are real, specific claims from Google's own safety team, though like the performance benchmarks, they currently rest on Google's internal testing and red-teaming rather than broad independent review, since the model isn't yet available for outside security researchers to test at scale.

Pricing: Introductory API Rates

Google has published specific pricing for Argon's API access: an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at a 95% discount off the standard input rate. Google states that after the introductory period ends, pricing rises to $4 per million input tokens and $20 per million output tokens. It's worth being clear that these are API developer prices, not a new price tag for existing Gemini app or Gemini Ultra subscriptions — if you use Gemini through your phone or a Windows PC, this pricing isn't something that changes your current subscription.

Availability: What's Actually Rolling Out Right Now

This is the single most important thing to understand about Argon today: almost nobody can use it yet. Access currently is limited to a set of trusted cyber defenders through Google's Fairwind Program, a program Google has described as being for governments and trusted partners handling cybersecurity-critical work. Google's announcement states broader availability will come "to developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers" — but gives no specific date for when that will happen.

What's Confirmed vs. What's Still a Claim

To be precise about the difference: Argon's benchmark scores, pricing, token limits, and the specific internal use cases (the libgav1 rewrite, the Wiz vulnerability discovery, the quantum optimization example) are Google's own published, specific figures, not vague marketing language, which gives them more weight than a typical hype announcement. Separately, Artificial Analysis independently measured Gemini 4 Argon at 53 on its Intelligence Index, matching GPT-6 Astra's score. Artificial Analysis reported that Argon's cost per task was about 60% of Astra's at the introductory pricing. This is useful third-party evidence, but it remains one independent evaluation rather than broad industry-wide verification.

What remains unverified by outside parties is how Argon performs in the hands of everyday developers and businesses outside Google's own testing and its small set of early partners, simply because that broader access hasn't happened yet. Framing from some coverage describing Argon as "retaking the benchmark lead" reflects Google's own reported scores against its own chosen comparison points, not a neutral, independently run head-to-head evaluation across all major labs. Until Argon reaches wider availability and independent researchers and ordinary users can test it directly, it's reasonable to treat the specific, detailed figures as credible evidence of real progress, while treating broader "best model available" framing as Google's marketing position rather than settled fact.

FAQ

Can I use Gemini 4 Argon right now?

Not through general public access. Google says Argon is currently rolling out to trusted cybersecurity defenders through its Fairwind Program. Broader access for paid API customers, developers, enterprises, and consumers is planned, but Google has not announced a specific public release date. Regular Gemini app users, including Google AI Ultra subscribers, don't have access yet, though Google has said they'll be among the first groups included once broader rollout begins.

Final Takeaway

Gemini 4 Argon is a detailed model launch with specific published benchmark results, internal use cases, and API pricing that can be checked against Google's announcement. However, most of the launch benchmarks and internal examples come from Google itself, while independent testing is still limited. At the same time, it's launching in an unusually restricted way, available only to a small group of cybersecurity partners for now, with no firm date for when the rest of us will get to try it. The honest summary: Google has shown real, specific evidence that Argon is a capable model, particularly for coding and cybersecurity defense, but broader, independent confirmation of how it performs in everyday use is still ahead, not behind us.

Sources

• Google: Gemini 4 Argon — Our Next Era of Frontier Intelligence
• Reuters: Google Announces Gemini 4 Flagship AI Model After Months of Delays
• Artificial Analysis: Gemini 4 Argon — Google Is Back as One of the Top Three Labs

Disclaimer

Performance figures for Gemini 4 Argon are primarily drawn from Google's own published benchmarks and internal use cases, with limited independent, third-party verification available at the time of writing, since the model has not yet reached broad public or developer availability. Figures and availability timelines may change as Google's rollout progresses; check Google's official announcements for the most current information.

Written by Mr. Tarsem Singh
Founder & Editor, Beinfora

This article was researched and written by Mr. Tarsem Singh to provide clear, useful technology information for readers