Google's Gemini 4 Carbon tops Artificial Analysis index, with cyber claims to check

Google announced Gemini 4 Carbon on October 28, 2026, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." The company says Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.

Carbon follows Gemini 4 Argon by two weeks. Google says Carbon "illustrates the pace at which we’re pushing the frontier forward." Two frontier models in a fortnight is a cadence, and also a good reason to read the tables before the press release.

The cyber numbers

The headline claim is security. Google says that, evaluated without guardrails on external cybersecurity benchmarks, Carbon scored 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching). The company describes this as "a firm lead." The post gives Carbon's scores and that phrase, not the runner-up's numbers, so the size of the lead is Google's word until someone lines up the columns.

There are supporting claims, and they vary in how checkable they are:

• Internally, Google says Carbon achieves more than a 2x uplift over Argon in success rate for generating secure code.
• Google's Chrome team reported that Carbon performed more than 10 times better than "another lab's large frontier model" at generating high-quality patches. The announcement does not name that model.
• Wiz, which Google calls its leading cloud security platform, observes Carbon leading on web penetration testing benchmarks, a core capability for its Scan for Good initiative.

A 10x figure against an unnamed competitor is a very large number attached to a very small amount of information. It may be entirely true. It is also impossible to reproduce from the post.

Behavior and safety figures

Google pairs the capability claims with a set of behavior metrics, all from its own announcement:

• On Terminal-Bench 4.0, successful cheating fell to 0.0% and attempts to cheat to 0.4%, while overall accuracy improved.
• In live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%.
• On Gray Swan's Indirect Prompt Injection Benchmark, Carbon's attack success rate is 0.0%, against Argon's 0.7%. Google says Carbon resisted every evaluated attack attempt.
• In internal testing, Carbon cuts unsupported claims by approximately 2x over Argon, reduces overreach by 26% and under-service by 29%.

A zero on a benchmark is a statement about that benchmark's attacks, not about the world's. A zero on a benchmark is a statement about that benchmark's attacks, not about the world's Likewise, 0.0% successful cheating is only as meaningful as the process that detects cheating. Several of these figures are internal, and the post does not say how the 80% drop in alerts was measured or what the baseline volume was.

Still, the direction is worth noting. Google is reporting not just how well Carbon performs but how often it misbehaves while doing so, and it is putting numbers on it. That is the right kind of claim to make, because it can be wrong in public.

Access, price and where Carbon is not first

Carbon is not broadly available. Starting this week, Google says it will make the model available to a subset of cyber defenders, then gradually phase in additional trusted partners through its Fairwind Program. The company also says it is "actively engaged in the U.S. government’s voluntary process for pre-release model testing." The announcement does not say which organizations are in the first group.

Pricing is public: the same introductory price as Argon, $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price.

And Carbon is not first at everything. On WebDev Arena it scored 1769, up 89 Elo points, which lands it at #4. Google says the gain comes from improved quality of output and artifacts in web development. Who holds the top three spots is not stated in the announcement.

What to watch

Four questions are open. Which lab's model did the Chrome team beat by 10x? When does Carbon reach standard developers, businesses and consumers? Who is in the first Fairwind group? And what do independent runs of SEC-bench Pro, ExploitGym and CWE-bench show once outsiders get hands on the model?

The Intelligence Index lead is a third-party score, though Google is the one reporting it. On Hacker News the announcement had 63 points and 26 comments. My take: show me the eval, not the vibe.