Dashboard

Gemini 4 Carbon: Google's Security-First Model

Google's new flagship leads on security benchmarks and is locked to vetted cyber defenders. The price signal and the injection numbers are the parts you can use.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
29 October 20261 min read

Google announced Gemini 4 Carbon on 28 October 2026, and called it its strongest model for cybersecurity and long horizon work. The catch sits in the same announcement: you almost certainly cannot call it. Access starts with a subset of cyber defenders and widens through a programme Google calls Fairwind. Developers, businesses and consumers get it later, with no date attached.

So the useful part of this release is not the benchmark crown. It is the price Google has set for frontier intelligence, and one safety number that changes how you should think about agents with tool access.

What Google says Gemini 4 Carbon is

Carbon is positioned for sustained agentic workflows and long, multi-step engineering tasks, rather than for chat. It builds on Gemini 4 Argon, which arrived about two weeks earlier and uses the same safety and security stack. If you missed that one, we covered what Argon actually shipped.

The numbers Google published

These come from Google's own announcement, and the security benchmark results are Google's own reporting rather than independent evaluation. Read them as vendor claims.

  • Artificial Analysis Intelligence Index v5.0: first place, score of 59.

  • SEC-bench Pro, finding vulnerabilities: 88.2 percent.

  • ExploitGym, exploitation: 54.8 percent.

  • CWE-bench, patching: 71.0 percent.

  • WebDev Arena: 1768 Elo, fourth place, up 89 points.

  • Gray Swan indirect prompt injection: 0.0 percent attack success rate.

  • Terminal-Bench 4.0: successful cheating reduced to 0.0 percent.

  • Internal testing: more than double the success rate for generating secure code, compared with Argon.

Google also reports that against Argon, unsupported claims drop by roughly half, overreach falls 26 percent, and under-service falls 29 percent.

The one number worth your attention

Indirect prompt injection went from 0.7 percent attack success on Argon to 0.0 percent on Carbon, on the Gray Swan benchmark. Indirect injection is the attack where the instructions are not typed by your user. They are hidden in the thing your agent reads: a web page, a calendar invite, a support ticket someone filed on purpose.

If your product reads untrusted text and then takes actions, that is the failure mode that empties a database or emails a customer list to a stranger. A vendor moving it toward zero on a public benchmark is a more meaningful improvement than an index score, because it is the number tied to a specific thing going wrong.

Treat 0.0 percent as a benchmark result, not a guarantee. It means the published attacks stopped working, not that injection is solved.

What Gemini 4 Carbon costs

Introductory pricing is 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input priced 95 percent below the input rate during that period. After the introductory period the rate becomes 4 dollars per million input and 20 dollars per million output.

That is the same introductory price as Argon, and the same doubling pattern. Argon ran the identical move, which we wrote up when Argon's price was set to step up. Two releases in a row have now used introductory pricing that doubles later, so if you are modelling costs on a launch rate, model the post-launch rate as well.

What to do about this today

  1. If you are shipping, nothing. You cannot get Carbon, and rewriting anything for a model you cannot call is wasted work.

  2. Take the price signal. Frontier intelligence is holding at 2 dollars in and 10 dollars out at launch, doubling afterwards. That is now the number to plan around.

  3. If you run an agent with tool access, use the injection figures as a prompt to audit your own exposure rather than to wait for Carbon.

  4. Keep your model choice boring. Our guide to picking a model without chasing every release applies here exactly: an announcement you cannot act on is news, not a migration plan.

Google's full post, with the benchmark tables, is on the Google blog. For the wider habit of filtering releases like this one, see how to keep up with AI news.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.