Google launches Gemini 4 Carbon: tops the index, but only cyber defenders get to try it

Google has launched Gemini 4 Carbon, a new frontier model the company calls its strongest yet. The model ranks first in analytics firm Artificial Analysis's updated index and, according to Google, sets new records in finding, exploiting and fixing security flaws. But for now, neither businesses, developers nor consumers can try it. Google is releasing Carbon first to a small group of cyber defenders.

The announcement came in a blog post on Wednesday 28 October at 17:00 UTC, written by Koray Kavukcuoglu, a senior vice president at Google DeepMind and Google's chief AI architect. Carbon arrives just two weeks after its predecessor, Gemini 4 Argon, according to Google.

Tops the index with 59 points

According to Google, Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. Google reports significant gains in coding and agentic work, meaning tasks where the model works independently across many steps.

The company writes that its own teams have seen Carbon work for days on complex engineering tasks, such as finding and patching subtle bugs across large codebases. On WebDev Arena, which measures web development, the model climbs 89 points to fourth place.

Records in security work

What Google highlights most is security. According to the company, Carbon scores 88.2 per cent on SEC-bench Pro, which measures the ability to find real vulnerabilities, 54.8 per cent on ExploitGym, which measures the ability to exploit them, and 71.0 per cent on CWE-bench, which measures whether the model can fix them. Google states that these tests were run without guardrails.

Google also cites internal results. Carbon is said to generate secure code with more than twice the success rate of Argon. The Chrome team reportedly found that Carbon produced high-quality security patches more than ten times better than a large frontier model from another AI lab, which Google does not name. Google-owned Wiz sees Carbon leading on benchmarks for web penetration testing, according to Google.

The company also reports better behaviour. Carbon never succeeded in cheating on the Terminal-Bench 4.0 test, attempted to in 0.4 per cent of cases, and resisted every attack in a test of prompt injection, where hidden instructions try to hijack the model. In Google's own coding environments, high-severity misalignment alerts dropped by almost 80 per cent. All of these figures are Google's own.

Only selected users get access

Google writes that it is taking a deliberate approach to the rollout and is taking part in the US government's voluntary process for pre-release model testing. This week, a subset of cyber defenders gets access. More trusted partners will then be phased in gradually through Google's Fairwind Program.

Developers, businesses and consumers will get Carbon through Google's products "as soon as possible", but Google gives no date. The post also says nothing about availability in Europe or Norway.

Pricing, however, is clear. Carbon launches at the same introductory price as Argon: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 per cent off.

What it means for leaders

Our assessment: Carbon shows that the largest AI companies now release frontier models only weeks apart, and that the ability to find and exploit security flaws is becoming a headline feature. Attackers can use the same capability, which is why Google is holding the model back. For Norwegian organisations this means two things. Security leaders should expect vulnerability discovery to become faster and cheaper, for defenders and attackers alike. And those already using Argon have a price basis to plan around, but no date for when Carbon will be available to them.

Sources and media

• Google blog: Gemini 4 Carbon: our next frontier model
• Google DeepMind on X: Gemini 4 Carbon announcement

Illustration: AI-generated with Grok Bot / hogby.ai