Google has a new frontier AI model, and most people cannot use it yet. On 30 September 2026, Google announced Gemini 4 Argon, a model built for long and difficult jobs in software, finance, law and cybersecurity.
The unusual part is who gets it first. Instead of opening it to everyone, Google is giving early access to a vetted group of cyber defenders. These are the teams that protect hospitals, energy networks, banks and government systems.
Google’s reason is simple. A model that is very good at finding security holes could also help attackers. So the company is starting small and widening access in steps.
Quick answer: Gemini 4 Argon is Google’s newest frontier AI model, announced on 30 September 2026 for long software, research and cybersecurity tasks. Right now it is rolling out to trusted security teams through Google’s Fairwind Program, and Google’s own teams use it too. Paid API customers and Google AI Ultra subscribers are next, and Google has not given a date for everyone else.
What Google announced
Google shared the news in an official blog post on Gemini 4 Argon written by Koray Kavukcuoglu, a senior leader at Google DeepMind. Here are the key facts in one place:
- Name and date: Gemini 4 Argon, announced on 30 September 2026.
- What it is for: long, demanding work in software engineering, business tasks such as legal and finance research, and cybersecurity defense.
- Longer answers: it can produce up to 1 million tokens in a single response, up from 64,000 before.
- First users: trusted cyber defenders in Google’s Fairwind Program.
- Next in line: paid API customers and Google AI Ultra subscribers.
- Launch price for developers: $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper.
What Gemini 4 Argon can do
Bigger jobs in one go
AI models read and write text in small pieces called tokens. A higher output limit means the model can write much longer results in one go, such as a full report or a large block of code. Google says this lets Argon reason more deeply on hard problems.
Google also shared examples from its own work. It says teams of Argon agents are working on moving C and C++ code to the Rust programming language, with projects of up to more than 800,000 lines. Another team of agents found memory savings that Google says will free up more than 300 tebibytes across its data centers once the changes roll out. If agents are new to you, our guide on what an AI agent is explains the idea in plain words.
Finding and fixing security flaws
According to the Google DeepMind page on Argon’s cyber skills, the model can do three main jobs:
- Find hidden weaknesses in large, complex code.
- Test a website from the outside, without seeing its source code, the way an attacker would.
- Write and check a fix once it finds a problem.
Google gave one real example from Wiz, the cloud security company it bought earlier this year. Wiz now runs inside Google Cloud. Wiz used Argon in its Scan for Good program, which finds risks in public infrastructure for free. Google says Argon found a critical flaw in healthcare software used by hospitals around the world, that the flaw exposed sensitive personal information, and that earlier top models had missed it. Google has not named the software, and this is a claim from Google’s own group, not an independent finding.
Why security teams get Gemini 4 Argon first
“Safely releasing frontier capabilities at this level requires a phased approach,” Google wrote in its announcement. In plain words, the company does not want the same skills that help defenders to land in the wrong hands on day one.
Gemini product lead Tulsee Doshi told CNBC that giving defenders first access builds confidence in the rollout and gets defensive tools to them sooner, as TechRepublic reported. Google also says it is taking part in the US government’s voluntary process for giving officials access to models before release.
What the Fairwind Program is
Fairwind started in early September and has more than 650 partners, according to SecurityWeek. The Fairwind Program page says access goes first to governments, national cyber agencies, operators of critical services such as healthcare, telecoms, energy and finance, major technology platforms and academic security labs.
The rules are strict. Google runs background checks on organizations that apply. Partners must use phishing resistant sign in and cannot pass their access to anyone else. Approved partners can use the tools for authorized threat simulation, reverse engineering and malware analysis, for defense and academic research only. Google does not allow harmful tasks such as creating malware. Google also says trusted defenders and its own teams get a version of Argon without the usual cyber guardrails.
Who can use it, and when
| Group | Access today? | What we know |
|---|---|---|
| Trusted cyber defenders (Fairwind) | Yes, rolling out now | Includes a version without the usual cyber guardrails |
| Paid API customers | Not yet, next in line | Launch price of $2 and $10 per million input and output tokens |
| Google AI Ultra subscribers | Not yet, next in line | Ultra is Google’s top paid plan, with two price levels |
| Everyone else | No date announced | Google says “as soon as possible” |
The fine print: benchmarks, doubts and price
Google says Argon scored 77.9% on DeepSWE v1.1, a test of real software engineering tasks. It also claims first place on AutomationBench with 51.3%, a tie for first place at 68% on a test of fixing security flaws, and 91.7% on a long video understanding test.
Independent testing tells a more mixed story. The Next Web reported that Argon scored 53 on the Artificial Analysis Intelligence Index. That matches GPT 6 Astra and sits one point above GPT 6.1 Sol. OpenAI shelved Astra on 28 September and released Sol the next day, Gizmodo reported. Anthropic’s Claude Opus 5.5 still leads with 58. On a factual knowledge test, Argon had a hallucination rate (how often it gives a wrong answer instead of saying it does not know) of 15%. GPT 6 Astra scored 51% and GPT 6.1 Sol scored 54%. But Argon also answered fewer questions correctly, 50% against 63% for GPT 6 Astra, because it more often admitted it did not know.
The same report noted a Bloomberg story that cited concerns inside Google about Argon’s coding skills. Google told Bloomberg it would be inaccurate to say the model underperforms in areas such as coding.
Cost needs a careful look too. The launch price is introductory, and Google lists a later standard price of $4 per million input tokens and $20 per million output tokens. In the independent tests, Argon used about 62,000 output tokens per task, against 27,000 for GPT 6 Astra. Even so, each task cost $1.99 with Argon at the launch price, against $3.26 for GPT 6 Astra. The standard price is double the launch price, so the same work could cost about $3.98 later. That is our own estimate, and it assumes the same token use. Real cost per job matters more than the price per token.
What it means for you and your business
For everyday users
You probably cannot use Argon in the Gemini app today. Google AI Ultra, which costs about $100 a month in the US for the entry plan and about $200 a month for the top plan, according to Google’s plans announcement, is in line for early access. Google has not given a date or said which Ultra plans will get it.
The bigger effect may be indirect. If defenders find and fix flaws faster, the apps and services you rely on, from hospital systems to banking, could get safer sooner. You can help by turning on automatic updates on your phone and computer, since fixes only protect you once you install them.
For businesses and developers
Here is a short checklist if you run a team:
- Security teams in critical sectors: check whether your organization fits the Fairwind priority list.
- Developers: plan for API access, but test Argon on your own real tasks before you switch.
- Finance teams: budget on cost per finished task, and plan for the higher standard price.
- Anyone building agents: keep human checks in place. Google says Argon is its most resilient model yet against indirect prompt injection, which is when hidden text in a web page or file tries to trick an AI into doing something you did not ask for.
What to watch next
- A firm date for paid API customers and AI Ultra subscribers.
- When the introductory price ends and the standard price begins.
- More independent tests on real work, not just benchmark scores.
- Whether the healthcare software maker shares details once a fix is out.
- Whether Argon reaches cheaper Google plans and the free Gemini app.
Frequently asked questions
What is Gemini 4 Argon?
It is Google’s new frontier AI model, which is the industry term for a model at the leading edge of what AI can do. Google announced it on 30 September 2026. It focuses on long software, research and cybersecurity tasks, and can write up to 1 million tokens in one response.
Can I use it in the Gemini app today?
Most people cannot. It is rolling out first to trusted security teams. Paid API customers and Google AI Ultra subscribers are next, but Google has not shared a date.
Why is Google giving it to security teams first?
Argon is very good at finding software flaws. Google wants defenders to use that skill first while it strengthens safeguards against misuse by attackers.
How much does it cost for developers?
For developers, the launch price is $2 per million input tokens and $10 per million output tokens. Google lists a later standard price of $4 and $20.
Is Argon better than ChatGPT?
It depends on the task. Google’s own tests show strong coding and security results. One independent index scored it level with GPT 6 Astra, which OpenAI has not released, and one point above GPT 6.1 Sol, which OpenAI has. Test it on your own work when you get access.
Final thoughts
Gemini 4 Argon follows a pattern that is growing in AI releases. Anthropic also limited its Claude Mythos Preview cyber model to selected partners in April. The most powerful models may now reach security teams before the public, and the benchmark claims still need real world proof.
Until Argon opens up, you have plenty of options. Compare the best AI tools for coding, try some free AI tools like ChatGPT, or start with the basics in our guide to what artificial intelligence is.