Technology

Google's Newest AI Model Is Good at Finding Software Weaknesses, and It Is Being Released Slowly

Martin HollowayPublished 2w ago3 min readBased on 5 sources
Reading level
Google's Newest AI Model Is Good at Finding Software Weaknesses, and It Is Being Released Slowly
source:blog.google

Google announced Gemini 4 Carbon on October 28, 2026. The company calls it its strongest AI model so far, with especially strong skills in cybersecurity and in long, multi-step tasks.

Carbon follows Gemini 4 Argon in the Gemini 4 family. Google said it launched Argon two weeks before the Carbon announcement. Reuters reported on September 30, 2026 that Alphabet had announced a new top-tier model to anchor the Gemini 4 family. On October 13, 2026, Reuters reported that Google had launched Argon for developers and businesses.

Google describes Argon as its most advanced model for real-world coding, office knowledge work, and cyber defense. Its Gemini models page carries the line "Built to solve your hardest problems." Google says developers, businesses, scientists, and consumers can all access it.

Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, a general test of AI ability, with a score of 59. On WebDev Arena, which tests how well models build websites, it ranks fourth with 1769. That is a gain of 89 Elo points. Elo is a rating system based on head-to-head comparisons.

Cybersecurity is the centre of the announcement. Google reports 88.2% on SEC-bench Pro for finding security flaws in software. It reports 54.8% on ExploitGym for exploiting those flaws, which means using them to break into a system. It reports 71.0% on CWE-bench for patching, or repairing, flaws. Google also says that on its internal tests, Carbon more than doubled the success rate for writing secure code compared with Argon.

Access to Carbon is being released in stages. Starting this week, Google plans to give it to a small group of cyber defenders, the security teams who protect computer systems. It then plans to widen access to more trusted partners through its Fairwind Program. Google also said it is taking part in the U.S. government's voluntary process for testing models before release.

For developers using Google's API, the interface that lets software talk to the model, introductory pricing is $2 per million input tokens and $10 per million output tokens. A token is a small piece of text, such as a word or part of one. Cached input tokens, which are text the system has already processed and stored, cost 95% less than regular input tokens.

In my view, the staged rollout matters as much as the benchmark table. A locksmith who can both build locks and pick them is useful to the people who need secure doors, and also dangerous in the wrong hands. A model that scores well at both finding and exploiting weaknesses works the same way, so deciding who gets it first is a product and policy choice as well as a technical one. The defensive results make the stronger case for the release. If a 71.0% patching score and the doubled secure-code rate hold up outside Google's test environment, known weaknesses could stay in live software for less time. Defenders are the group most likely to turn that into fewer exposed systems, which is why their early access is the detail to watch.

The next signals will come from outside Google's own measurements. Independent tests of the cyber benchmarks, the pace of the Fairwind Program, and the effect of government pre-release testing on the timing of wider access will all be worth following.