This article contains affiliate links. As an Amazon Associate I earn from qualifying purchases.
On September 30, 2026 (US time), Google announced its new top-tier model, Gemini 4 Argon. Gemini 3 came out in November 2025, so this is the first generational step in roughly ten months. What stood out most this time, though, was not the performance numbers but the way the model is being handed out. It has been announced, yet neither ordinary users nor developers can touch it. The first to receive it were a limited set of organizations on the defending side of cybersecurity.
Before going further, a note on where I stand. The articles on this blog are written with Anthropic’s Claude, and I also build and distribute AI apps that run locally. When reading what follows about a Google model, keep in mind that it is written by someone who uses a competitor’s tools. All figures come from either the official announcement or independent evaluation sites, and I will make clear which is which.
A Gap Filled With Flash Models
Since Gemini 3, Google’s new releases had been 3.1 Pro Preview, followed by a run of the lighter, faster Flash line: 3.8 Flash, 3.8 Flash Cyber, and 3.8 Live. By the count of the independent evaluation site Artificial Analysis, this is the first model above the Flash class in seven months. According to Bloomberg, Gemini 3.5 Pro, which had been promised for June, was ultimately dropped; resources were shifted to 3.8 Flash, and the company skipped a generation to arrive at 4.
In the meantime, OpenAI shipped its GPT-6 line and Anthropic its 5.5 line, and Google looked like it had stepped back from the race at the top. The announcement was written by Koray Kavukcuoglu, SVP at Google DeepMind, who calls Argon the “next era of frontier intelligence.” That said, he had already told reporters last week that Google wanted Gemini 4 out well before the end of the year, so the timing is on schedule.
One Million Tokens in a Single Output
The clearest change in the specifications is the output limit, raised from 64,000 tokens to 1 million. The input context window is also 1 million tokens. In short, long tasks can now run from thinking through to finished writing in one pass, without being chopped up midway. The API also gained a feature called Long Decode Continuation, which pauses a long response and resumes it in the next call, so a job can run to completion without hitting timeouts.
A few of the numbers Google cites: 77.9% on DeepSWE v1.1, which measures long, real-world software engineering tasks; first place on the Vals Index, which weights finance, legal, tax, and coding work by economic value; first place at 51.3% on AutomationBench, Zapier’s business-automation benchmark; and 91.7% on LVBench, which tests understanding of long videos. Inside the company, thousands of Google employees are reportedly already using it for everyday development and research.
Two internal examples stand out. One is memory optimization in the data centers. A group of Argon agents analyzed profiling data from across Google’s fleet, found and applied optimizations on their own, and freed more than 300 TiB of memory once rolled out (the eventual estimate is 500 TiB to 1 PiB). The other is migrating C/C++ to Rust. For the libgav1 video decoder, Argon agents took an existing Rust port, rewrote its 32,000 lines of SIMD code in safe Rust so the compiler would vectorize it automatically, and ended up 2.7 times faster than the Rust port. A migration of Fuchsia’s Zircon kernel, at more than 800K lines, is said to be in progress as well.
The image above is an artist’s impression. The actual charts and tables are on Google’s announcement page, linked at the end of this article.
Defenders Get It First
This is the heart of the announcement. Argon is initially going only to trusted cyber defenders, through Google DeepMind’s Fairwind Program. Fairwind gives early access to new models to governments and national cyber authorities, operators of critical infrastructure such as hospitals, telecoms, and power grids, and the makers of foundational software that many people depend on. More than 650 organizations are said to be participating. Applicants are vetted for their track record and ethics, and reselling or sharing access is prohibited.
Within this program, moreover, Argon is provided with its cyber guardrails removed. It finds vulnerabilities, confirms they can actually be exploited, and patches them, all autonomously. In the hands of defenders that is an inspection tool; in the hands of attackers it is an attack tool outright. Google itself writes in the Fairwind description that the same capabilities become an equally powerful threat in the wrong hands. So the reason for delaying public release is not that the model is weak, but that it is too strong in one particular direction.
Concrete examples of that capability are already out. The security company Wiz used Argon in its Scan for Good initiative, which inspects public infrastructure for free, and found a critical vulnerability exposing personal information in healthcare software used by hospitals worldwide. According to Google, previous frontier models had missed it. On CWE-bench v1, which measures the ability to fix vulnerabilities, Argon scored 68%, tied for first with Grok 4.7 and GPT-6 Astra (Claude Opus 5.5 scored 67%). On Wiz’s penetration-testing benchmark, which probes live web systems without source code, Argon scored 70.9% against 58.2% for its predecessor, 3.8 Flash Cyber.
Google also lists four safeguards it is firming up before broader release: refusing misuse related to chemical, biological, radiological, and nuclear weapons; resistance to indirect prompt injection (attacks that hijack a model through instructions hidden in documents or web pages); monitoring the model’s reasoning and actions to detect when it is going beyond the user’s intent, and stopping it when necessary; and running high-risk training and evaluation in sealed environments. On the monitoring point there is a notable caveat: findings from the monitors are not fed back into training, because doing so could teach the model to evade them. Google is also taking part in the US government’s voluntary process for pre-release model access.
What Shows Up Outside the Benchmarks
On the day of the announcement, Bloomberg ran a story from a different angle. Some Google employees, it reported, see Argon as strong on benchmarks but less impressive in actual work, with moments where it stumbles on coding. Google replied that it would be inaccurate to say the model underperforms in areas like coding, and one employee said there is a “large consensus” inside the company that it is a frontier model. Reports of mixed feelings inside a company are nothing unusual, but coming after ten months of Flash-only releases, this top-tier model is drawing a different kind of attention.
Third-party numbers are beginning to appear. On the Artificial Analysis composite index, Argon scores 53 at its highest reasoning setting, level with GPT-6 Astra and one point above GPT-6.1 Sol. That is 23 points above Google’s previous Pro-class model (3.1 Pro Preview, at 30), and the site says Google is back among the top three labs in intelligence.
Break it down by category, though, and the picture is uneven. On Terminal Bench 4, which covers work done in a terminal, Argon scores 57%, a 53-point jump from 3.1 Pro Preview, but still short of Claude Sonnet 5.5 at 64%, Claude Opus 5.5 at 60%, and GPT-6 Astra at 59%. On AutomationBench-AA for business automation, it leads at 78%. Another interesting figure is the hallucination rate (confidently stating things that are not true): on the site’s AA-Omniscience test, Argon comes in at 15%, the lowest of any model scoring 45 or more on the composite index. GPT-6 Astra is at 51% and GPT-6.1 Sol at 54%. The trade-off is accuracy, at 50%, down 5 points from 3.1 Pro Preview and 13 points below GPT-6 Astra’s 63%. The model has been tuned toward admitting what it does not know, and correct answers dipped a little as a result.
A word on cost, too. Artificial Analysis puts the cost per task at $1.99 at the discounted price, about 60% of GPT-6 Astra’s $3.26, but that is because the unit price is lower, not because fewer tokens are used. Argon averages 62,000 output tokens per task, more than double GPT-6 Astra’s 27,000. Once the discount ends, the cost per task rises to $3.98, slightly above GPT-6 Astra.
When and at What Price
After the Fairwind defenders, the next in line are paid API customers and Google AI Ultra subscribers, followed by developers, enterprises, and general users. No specific dates have been given.
Pricing at launch is discounted: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. The end date of the discount is undecided; afterward, the price goes to $4 for input and $20 for output. Since the output limit has grown to 1 million tokens, the amount consumed in a single call can grow accordingly, so for long delegated tasks the total volume will matter more than the unit price.
To sum up, Gemini 4 Argon moves ahead on three fronts: finishing long tasks in a single pass, ranking high on evaluations of professional work such as finance and law, and finding and fixing vulnerabilities. Because the third of these is so strong, the model has entered the world by going to defenders alone first. It falls short of rivals on some benchmarks, and cautious voices have been reported inside the company. The sensible thing is to wait until it is available to use and then see how far it can be trusted with long jobs.
Related products
You can find books and products related to this article on Amazon.
If you buy through the links above, this blog may earn a commission.
Sources
Google official blog (Koray Kavukcuoglu, September 30, 2026): Gemini 4 Argon: our next era of frontier intelligence
Google Japan Blog (Japanese edition): Gemini 4 Argon：フロンティア知能の新時代
Google DeepMind: Fairwind Program
Artificial Analysis (September 30, 2026): Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved
The Next Web (Ana Maria Constantin, September 30, 2026; includes a summary of Bloomberg’s reporting on internal skepticism): Google unveils Gemini 4 Argon, and cyber defenders get it first
CNBC (September 30, 2026): Google rolls out Gemini 4 Argon, its most advanced AI model
コメント