Google Unveils Gemini 4 Carbon, a Cyber-Focused AI Model That Tops New Index

Google has introduced Gemini 4 Carbon, a frontier AI model built for long-running software engineering and cybersecurity work. It has taken first place on Artificial Analysis’ updated Intelligence Index, but this is not an immediate public launch. Google says selected cyber defenders will get access first, with a wider release to developers, businesses and consumers to follow after further testing.

That limited rollout is central to the story. A model able to find, validate and patch software vulnerabilities could help defenders close gaps faster. The same skills could also be useful to people trying to exploit them. In its 28 October announcement, Google says it is evaluating safeguards and participating in a voluntary US government process for pre-release model testing before broadening access.

What Gemini 4 Carbon can do

Carbon is designed to persist through complex, multi-step tasks rather than answer a single prompt and stop. Google describes internal runs lasting days as the model investigates subtle bugs across large codebases, identifies a root cause and attempts a fix. The company says it also improved performance in coding, scientific tasks, full-stack web development and multimodal understanding. These examples are useful indications of Google’s intended uses, but they are not a guarantee that every user can leave it to work unattended.

The most striking numbers concern cybersecurity. Google reports an 88.2% result on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for validating whether flaws can be exploited, and 71.0% on CWE-bench for producing patches. Those are different tests with different scoring rules. They should not be collapsed into a single claim that Carbon finds or fixes a given percentage of security problems in the wild.

Google also says Carbon more than doubled its success rate for generating secure code over Gemini 4 Argon on an internal benchmark. Its Chrome team reported substantially better patch generation than a rival frontier model, while Wiz found strong results in web penetration testing. Because those comparisons are presented by Google and the full conditions are not laid out in the announcement, security teams will still need to test the model against their own code, tools and review procedures.

Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, placing it first on the evaluator’s current leaderboard. Artificial Analysis says the index combines nine evaluations covering areas such as mathematics, science, coding, reasoning and agentic work. The next listed model, Claude Opus 5.5 at its Max setting, scores 58. This is an independent overall capability ranking, not an independent certification that Carbon is the safest or best choice for every defensive security task.

The index itself was recently revised, so comparisons with scores from older versions need care. A one-point lead is noteworthy, but procurement teams should also examine cost, speed, reliability, guardrails and task-specific results. A high composite score does not show how a model behaves when it encounters a private repository, uncertain evidence or a change that could break production software.

Google says Carbon builds on the safeguards introduced with Argon, including misuse defences, prompt-injection resistance and monitoring for unsafe behaviour. It claims fewer unsupported statements and fewer serious alignment alerts in internal coding tests. Those claims deserve scrutiny as usage expands. The practical safety question is whether a powerful model reliably stays within the defender’s authorised scope while handling unfamiliar systems and ambiguous instructions.

TechBooky previously covered Gemini 4 Argon’s restricted debut and its later expansion to paid users and developers. Carbon follows a similar staged path, but Google is initially targeting a subset of cyber defenders through its Fairwind programme. It has not given a firm date for general access. Users should therefore not assume Carbon is already available in the Gemini app or through the standard API.

Google states an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input priced at a 95% discount. That is a proposed price for access when it is offered, not evidence that everyone can buy it today. For long-running agents, the final bill will depend on how much context they read, how many attempts they make and whether a human has to repeat or repair their work.

Carbon’s launch raises a question beyond the leaderboard. If AI systems become genuinely effective at discovering and fixing weaknesses, who gets them first, and how quickly can defenders turn a finding into a tested patch? Google’s initial answer is controlled access and more safety evaluation. The evidence that will matter next is what authorised security teams can achieve with the model on real systems, and whether those gains hold when access widens.