Gemini 4 Carbon Tops the Artificial Analysis Index and Goes to Cyber Defenders First

Google has released its second frontier model in two weeks. Gemini 4 Carbon, announced by Sundar Pichai on LinkedIn on 28 October, follows Gemini 4 Argon and ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The first organisations to receive it will be a subset of cyber defenders vetted by Google, starting this week.

Where Carbon leads

Security carries most of Google’s case for the new model. Carbon scored 88.2% on SEC-bench Pro, a test of how well a model finds vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for writing patches. Google’s Chrome team reported that Carbon generated high-quality patches more than ten times better than a large frontier model from another lab, without naming that model or the method behind the comparison. Anthropic made security its showcase in a similar way in April, when its Mythos model found 271 Firefox security bugs.

Google’s announcement also reports changes in how the model behaves during long coding sessions. In live internal environments, high-severity misalignment alerts fell by almost 80% compared with Argon, and claims the model could not support were roughly halved. Carbon will launch at Argon’s introductory price of $2 per million input tokens and $10 per million output tokens, which rises to $4 and $20 once the introductory period ends.

Who gets it first

Google is distributing Carbon through the Fairwind Program, its early-access scheme for defensive security work. The programme gives priority to governments, core technology platforms and critical infrastructure operators, and it lists financial networks among the infrastructure it covers. Partners go through background checks and may use the model only for defensive tasks such as authorised threat simulation, reverse engineering and malware analysis.

For banks, payment companies and other fintech firms, applications go through their Google Cloud account team. Mistral took a comparable route on 6 October, when Mistral Large 4 arrived with stronger cyber capabilities and lighter moderation reserved for security specialists, approved partners and government agencies.

Testing before release

Google, which signed a pre-release testing agreement with the Commerce Department’s Center for AI Standards and Innovation (CAISI) in May, says Carbon is going through the US government’s voluntary testing process, although the announcement does not name the agency. Those voluntary model reviews rely on labs choosing to submit their models, since the government cannot require submissions or block a release.

Developers, businesses and consumers will get Carbon once Google has gathered feedback from partners and refined its guardrails, and the company has given no date. Until then, the most capable model Google has built will spend its first weeks searching other organisations’ code for flaws, and the first verdict on it will come from security teams whose job is to break things before attackers do.

Editor’s note

Every piece goes through careful verification, but mistakes can happen. Readers who spot an error or have additional information can write to rosalia@thebrightminded.com.