AI
Google's new AI finds security holes in software. Why you can't use it yet
Gemini 4 Argon, Google’s newest AI model, has no public release date. Google announced it on 30 September 2026 and gave it first to approved cyber defenders, because it’s very good at finding security holes in software. Google AI Ultra subscribers and companies that build on Gemini come next. The free Gemini app and Gemini in Gmail stay as they are for now.
Published 30 September 2026, updated 13 October 2026.
Why defenders get it first
Argon finds holes in software (mistakes in the code that let a stranger in), checks them and writes the fix. The security company Wiz used it to find a critical hole in healthcare software used by hospitals around the world, one that earlier models had missed.
Attackers could use it the same way, so Google lets defenders close holes first. OpenAI did the same with GPT-6 Astra on 3 September.
Who has it, and who's next
| Who | When |
|---|---|
| Fairwind security teams and Google’s own teams | Now, since 30 September 2026 |
| Companies that pay to build software on Gemini | Next, no date |
| Google AI Ultra subscribers | Next, no date |
| Everyone else, including the free Gemini app | Not announced |
Sources: Google, 30 September 2026, Google on Fairwind, 2 September 2026, Google DeepMind Fairwind page, Security Magazine, 3 September 2026.
The API price
Google has announced $2 per million tokens in and $10 out at first, then $4 and $20. It isn’t in the Gemini API model list yet (13 October 2026).
Sources: Google, 30 September 2026, Gemini API models (checked 13 October 2026).
Until it arrives
The ChatGPT, Claude and Gemini plans you can buy today handle the everyday work, and our comparison helps you pick one. Attackers will get tools like this too, so turn on two-step login, starting with your email, and keep your website and its plugins updated.
If nobody checks your website’s security today, that’s the gap to close. It’s part of how we look after our clients’ websites and apps.
Frequently asked questions
Short answers
Google hasn’t given one. In remarks reported on 24 September 2026, Koray Kavukcuoglu, Google’s Chief AI Architect, said he hoped for a release "much earlier" than the end of 2026 (9to5Google).
A closed program, opened on 2 September 2026, that gives governments, Google Cloud customers and security partners early access to Google’s cyber defence AI. It has more than 650 partners, and Google checks the background of each one that applies.
Guardrails are the limits that make an AI refuse harmful requests. Approved defenders get Argon without its cyber guardrails so they can use all of its security skills. For everyone else, Google says it’s built to refuse harmful cyber requests.
Didn't find your answer?
Ask us directly. You talk to the people who build it.
15 min on Google Meet
Found a mistake, or a story we should cover? Email contact@pc113studio.no. If something is wrong, we fix it and note the change in the post.
New posts by email, once a week
One short email with the week’s posts on the topics you choose. Leave whenever you like.
More from the blog
Request your free website demoBefore you pay a single cent.
We design and build a demo of your new website and send you the link.
See how your new website could look.
Love it? We agree the price and finish it. If not, it costs you nothing.
- Your first version in 7 working days
- No commitment on your side
- Free of charge
We take on a limited number of demos each month.
Tell us what you need
A few lines is enough
Website, ecommerce, app, campaign or AI assistant, or an idea still taking shape. We reply personally by email.
Fill in the form, or reach us directly:
Visit us in Sandvika
Elias Smiths vei 25, 1337 Sandvika
Opening hours
- Monday–Friday
- 09:00–19:00
- Saturday
- 11:00–16:00
- Sunday
- Closed
Open in Google Maps© OpenStreetMap