Gemini 4 Argon was Google’s newest frontier model for only two weeks. The company has now announced the next one, Gemini 4 Carbon.
Google says Carbon beats Claude Opus 5.5 in several benchmark categories, and it’s leaning hardest on coding and cybersecurity.
“Building on Gemini 4 Argon , which was launched two weeks ago, Carbon illustrates the pace at which we’re pushing the frontier forward,” Google said in a blog post.
A narrow lead on the general index
The headline figure is from the AA Index v5.0, a general AI benchmark. Google says Carbon scored 59 there, which puts it first, slightly ahead of Claude Opus 5.5 and GPT-6 Astra.
A slight lead says little by itself. For months, OpenAI and Anthropic have gone back and forth, each claiming wins in different categories. A small margin on a general index usually flips when the next model comes out.
Google also says Carbon won almost every category it tested, including agentic coding, computer use, knowledge work, science and math. Google ran those benchmarks itself, so independent testing will have to confirm the results.
Security is where Carbon pulls ahead
According to Google, cybersecurity is the category where Carbon beats other models by fairly wide margins.
The company says Carbon is good at finding vulnerabilities, ranking them by priority, and fixing them at the root cause. That’s harder than patching over the symptom. Google also says Carbon is twice as likely as Argon to generate secure code.
That secure-code figure is the one developers should watch. Coding assistants often produce code that runs but has holes in it. If Carbon really halves that rate compared with the previous model outside Google’s own tests, that would be a measurable improvement.
Honesty is on the scorecard too
Google says Carbon improves frontier model safety and security in several ways. It takes fewer shortcuts, exploits fewer loopholes, holds up better against prompt injection, and is more transparent and honest. One test covers “unsupported claims,” which tracks how often a model states things it can’t support.
The company also says Carbon collaborates better and is better at “following through without overstepping.” That phrase is aimed at agentic tools that go further than the user asked.
Not available yet
Carbon is getting the same rollout as Google’s earlier frontier models. A handful of cybersecurity defense companies receive it first for pre-release testing. After that, Google will give more trusted partners access through its Fairwind Program, and then release it widely.
Argon went through the same process and is now in the Gemini app and AI search mode. For most people, that’s as close as they’ll get to Google’s latest model until Carbon completes the same cycle.