Google's Gemini 4 Carbon Targets Cyber Defence, Not Consumers
Google has unveiled Gemini 4 Carbon, its latest frontier model, aimed at finding and fixing security flaws rather than at the general chat market. The company says it delivers state of the art results on cybersecurity tasks, including vulnerability discovery, exploitation and software patching, and generates secure code at more than twice the success rate of Gemini 4 Argon on internal benchmarks.
The model arrives two weeks after Argon, a pace that Google attributes to its researchers and models working in tandem. On the revised version of the Artificial Analysis Intelligence Index, Google says Carbon ranks first with a score of 59, and it reports 88.2 per cent on SEC-bench Pro for finding vulnerabilities, 54.8 per cent on ExploitGym for exploitation and 71 per cent on CWE-bench for patching.
And the framing matters. A model that can exploit flaws is also a model that can defend against them, which is why Google is keeping this one away from the public at launch and handing it to security teams first. Defenders get the tool before attackers do, at least in principle, and the staged rollout is meant to keep it that way.
Access runs through the Fairwind Program, which gives selected cyber defenders early use, with trusted partners added gradually while Google collects feedback and adjusts its guardrails. Developers, businesses and consumers will get it later, through the company's product surfaces, on a timetable the company has not fixed. Google also says it is taking part in the United States government's voluntary pre release testing process.
The price is set to match Argon's introductory rate of USD 2 per million input tokens and USD 10 per million output tokens, with cached input tokens discounted by 95 per cent. The introductory rate is a fraction of what the company has said standard pricing will be once it expires.
But the benchmark claims deserve reading with care. They come from Google, they are measured on internal tests and on third party leaderboards the company chose to cite, and the headline comparison is with its own previous model rather than with a named rival. The company does say Carbon cut unsupported claims by roughly half against Argon and reduced task overreach by 26 per cent, which speaks to reliability as much as to capability, and it is why the headline numbers should be treated as a starting point rather than a verdict.
The launch is also a statement about where the frontier competition now sits. Coding and security work are the tasks enterprises pay for, and they are the tasks where an agent can be measured on whether the job got done. Google, OpenAI and Anthropic are all pushing in the same direction, and each release sets the bar the others have to clear.
For security teams in Asia, the practical question is access rather than benchmarks. A model that finds vulnerabilities faster is only useful if the defenders using it can act on the findings, and early cohorts will show whether Carbon shortens the time between a flaw being found and a patch shipping.
Google launched Gemini 4 Carbon, a frontier model focused on cybersecurity and long running engineering work
Google says it ranks first on the Artificial Analysis Intelligence Index at 59
Access starts with selected cyber defenders through the Fairwind Program, with wider release later
Introductory pricing is USD 2 per million input tokens and USD 10 per million output tokens
Carbon is claimed to generate secure code at more than twice the rate of Gemini 4 Argon
