Google announced Gemini 4 Argon twice. The first time, on September 30, it was a limited release: access started with a group of trusted cyber defenders, with the public rollout promised “soon” and no date. The second time, on October 13, the company said the model was available to developers, businesses, researchers and consumers. This article is based on Google’s own pages for both announcements, the original DeepMind post, the launch post and the Google Cloud availability note. Nearly every performance number in them is Google’s own, and we label it that way.
Two announcements, two weeks apart
The staged approach is the real story behind the word “finally.” In the September post, Koray Kavukcuoglu, Google’s chief AI architect, described Argon as a frontier model for “real-world coding, enterprise knowledge work, and cyber defense,” and said releasing capabilities at this level “requires a phased approach.” The first group was defenders in DeepMind’s Fairwind Program, with Google gathering feedback and tightening guardrails before the first wider release to paid API customers and Google AI Ultra subscribers.
By October 13 the list was longer. Google says Argon is available through the Gemini API in AI Studio and in Android Studio, in Google Antigravity, in the Gemini app and AI Mode in Search for AI Ultra and Pro subscribers, and for businesses through Gemini Enterprise. On Google Cloud the model ID is gemini-4-argon.
What Google says it is good at
Google’s pitch is sustained, multi-step work: coding, finance, legal research, cybersecurity and science. The company says Argon works longer on hard tasks, writes better, repeats itself less and hedges less than earlier Gemini models. It also lists stronger spatial reasoning, mathematical precision and visual coding.
On size, the posts differ in wording. The September post says the output limit grew to 1 million tokens from 64,000. The October post speaks of a “1M unified input-output token limit.” The Cloud page gives no figure and points to the model card. If your project depends on the exact limits, read the model card before you plan around either phrasing.
How to read the benchmark list
Google reports first place on the Vals Index, which covers finance, coding, legal and tax work, and a lead on Harvey’s legal benchmark. It reports 77.9% on DeepSWE v1.1, which it calls a new state of the art for long software tasks, 51.3% on Zapier’s AutomationBench, 91.7% on LVBench for long video, and a tie for first on CWE-bench v1 at 68% for fixing vulnerabilities. On embodied reasoning, it cites 87% on ERQA.
These are all vendor-reported. Several come from third-party benchmarks, which is better than internal tests, but we have not seen independent replications, and “leading” on a benchmark is not the same as being better on your own task. The more useful numbers may be Google’s internal ones: a rise of more than 35% in agentic code submissions and nearly a third fewer rollbacks in Antigravity, and faster session preparation in its Agent Development Kit. They describe Google’s own engineering work, which will not match yours.
What it costs: $2 now, $4 after November 30
Pricing is where the planning matters. Until November 30, 2026, Google charges $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After that date the price is $4 and $20. In plain terms, it doubles.
That makes the next seven weeks a testing window, not a price. If you are building on Argon, estimate your costs at the November rate, not the introductory one. This echoes a point we made in our piece on why cheaper AI can still push compute demand up: a low unit price encourages heavy use, and heavy use is what drives the bill.
Google’s own proof points
The September post includes examples from inside Google. Agents analyzed profiling data across data centers and found memory optimizations that Google expects to free more than 300 TiB once rolled out. Others are migrating C and C++ code to Rust, from small libraries to over 800,000 lines in the Fuchsia Zircon kernel. In one case, a video decoder’s 32,000 lines of hand-tuned code were replaced with safe Rust, and Google says the result runs 2.7 times faster than the existing Rust port with identical output.
The October post adds claims in other fields: faster computation in a quantum chemistry package, a version of a quantum error decoder with nearly doubled throughput, and enzyme designs with lab-confirmed activity. These are interesting, but they come from Google and its partners and have not been independently reviewed here.
The cyber question: Argon without guardrails
The most unusual decision concerns security. Argon is trained to find, validate and patch serious software flaws. Google says it will provide a version without cyber guardrails to vetted defenders and its own teams, through an application-based Fairwind Program, expanding to hundreds of additional partners. In an early demonstration, security firm Wiz reportedly used it to find a critical flaw exposing personal data in healthcare software, which Google says earlier frontier models missed.
That is a sensible rule in principle, since defenders need the full capability. It also puts the weight on who counts as a defender, a decision outside the reach of anyone reading a blog post. Ordinary users get the guarded version.
Safety claims worth reading slowly
Google calls this its most comprehensive safety stack so far and says it tested under its Frontier Safety Framework. Its figures include attack success cut to 0.7% across 15 adaptive attempts on a prompt-injection benchmark from Gray Swan, shortcut attempts on coding benchmarks down to 2%, and external red teams unlocking under 1% of tested biology sub-tasks. It also describes a real-time monitor that reads the model’s reasoning and actions and says it avoided training against the model’s reasoning to keep it monitorable.
Each number is a lab result under specific conditions. They are a reason to look at the model card, not a guarantee for a deployment with real data. Agents that read untrusted web pages and documents still need their own protections.
What to test before you switch
If you already use Gemini, the sensible move is a short, concrete trial. Take three or four real tasks, run them on Argon at the November price, and measure quality, time and cost against what you use now. Check the token limits in the model card. Note that the Gemini 4 family is described as just beginning, with more models promised, so a smaller and cheaper sibling may arrive. And if your main interest is business automation, see how Argon fits into Google’s Gemini agent, which can route work between models.
Editor’s note: This article was researched, drafted and revised with the help of Claude, an AI assistant made by Anthropic, working from official pages and the reporting linked in the text. Nicolae Panait, editor and owner of Webitlio, chose the topic, reviewed the facts and made the final editorial decisions.
