Google announces Gemini 4 Carbon, its new flagship model, initially reserved for cybersecurity specialists

Google announced this Wednesday, October 28, around 6 p.m. Paris time, Gemini 4 Carbon, presented as its most powerful model, with an unusual emphasis: cybersecurity and long-horizon software engineering tasks. It arrives just two weeks after Gemini 4 Argon. But unlike Argon, Carbon is not yet accessible to developers or the general public.

Two days ago, we wrote that a “Carbon” version remained at the rumor stage, based solely on a Business Insider article. Google has now officially confirmed it.

A restricted-access launch

This is the most important point for anyone wanting to try it. According to the post by Koray Kavukcuoglu, Google’s head of AI architecture, Carbon will first open starting this week to a group of cybersecurity defenders, then gradually to other trusted partners via the Fairwind program, the same one that received Argon early. Google says it wants to gather feedback and adjust its guardrails before opening up to developers, businesses, and individuals, “as soon as possible,” with no date given.

Google also notes it is participating in the voluntary pre-release model testing process established by the U.S. government.

Pricing is already set: $2 per million input tokens and $10 per million output tokens during a launch period, same as Argon, with a 95% discount on cached inputs. Afterward, the rate will rise to $4 and $20, double the launch price. Google does not specify how long this launch period lasts.

What Google is highlighting

According to Google, Carbon takes 1st place on Artificial Analysis’s Intelligence Index v5 with 59 points. To put that number in context: in the v5 ranking published by Artificial Analysis, which we analyzed earlier today, Claude Opus 5.5 led with 58 points and Gemini 4 Argon was at 50. The announced leap would therefore be significant, but it is a one-point margin at the top, on an index that was just overhauled.

In cybersecurity, Google cites 88.2% on SEC-bench Pro (finding real vulnerabilities), 54.8% on ExploitGym (verifying that a vulnerability is exploitable), and 71.0% on CWE-bench (fixing the root cause), tests it says were run without guardrails. The Chrome team reportedly achieved patches “more than ten times” better than with another lab’s large model, unnamed. On the web development side, Carbon would gain 89 points to rank 4th on WebDev Arena (1769 points).

Google finally emphasizes the model’s behavior: fewer shortcuts and less test-gaming (0.0% successful cheating on Terminal-Bench 4.0), roughly half as many unsupported claims as Argon according to its internal tests, and no successful attacks via indirect prompt injection on Gray Swan’s benchmark. Internally, its teams reportedly saw it work for days on complex engineering tasks.

The limitations

All of these numbers come from Google. None have yet been independently verified, since almost no one has access to the model. Several rely on internal evaluations, and the cited rankings move quickly. The post also gives no context window size, supported modalities, or timeline for broader availability.