Gemini 4 Carbon: Google's Security-First AI Model Tops the Rankings

By TaazaPulse News Desk · Published: 29 October 2026 · TaazaPulse International Desk

Google DeepMind has unveiled Gemini 4 Carbon, a new frontier AI model built for two of the hardest jobs in computing — cybersecurity and long-horizon software engineering — and, unusually, has decided that most people cannot have it yet. Announced on 28 October, Carbon tops a leading independent intelligence ranking on the day of its launch, but its rollout is deliberately gated: it reaches vetted cyber defenders first, through a programme Google calls Fairwind, precisely because the abilities that make it useful for defence also make it powerful for attack.

What Google announced

The announcement, published under the name of Koray Kavukcuoglu, Google DeepMind's chief technology officer, positions Carbon as the second release in the Gemini 4 family, arriving two weeks after Gemini 4 Argon. Where Argon was the general-purpose flagship, Carbon is a specialist frontier model: trained and evaluated for finding and fixing security weaknesses in real codebases and for sustaining long, multi-step coding work — the kind of agentic tasks where a model must plan, act, check its own output and continue for hours rather than seconds.

Google's headline claim is the top spot on the Artificial Analysis Intelligence Index v5.0, where Carbon scores 59 — ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 "Astra" at 57, with Argon at 50. On the WebDev Arena leaderboard, which ranks models on building web applications, Carbon places fourth with an Elo of 1,769, a jump of 89 points over its predecessor.

The security benchmarks — and the catch

The numbers that define Carbon are its security evaluations, which Google published for the model's unguarded checkpoint: 88.2 per cent on SEC-bench Pro, 54.8 per cent on ExploitGym and 71.0 per cent on CWE-bench — benchmarks that measure, in plain terms, how well a model can find real classes of software weakness and reason about exploiting or fixing them. Those are capabilities with two faces. In a defender's hands, a model that can audit a codebase and surface exploitable flaws before criminals do is a force multiplier. In the wrong hands, the same model shortens the path from a published vulnerability to a working exploit.

That duality explains the launch strategy. Carbon is being released first through the Fairwind Program to vetted defenders — security teams and researchers whose identity and purpose Google can check — rather than as a general API product. Google also says the model underwent voluntary pre-release testing with the US government, part of an emerging pattern in which frontier labs submit their most capable systems for external scrutiny before wide release. The phased rollout means independent verification of the benchmark claims will arrive gradually, a caveat worth holding: launch-day leaderboards are the vendor's own framing, and the index providers' ongoing testing will confirm or adjust the picture.

Reading the benchmarks in plain language

Benchmark names mean little outside the field, so it is worth translating Google's three headline numbers. CWE-bench tests whether a model can recognise entries from the standard catalogue of software weakness types — the classic bugs, from injection flaws to broken access control, that real breaches are made of. SEC-bench Pro raises the difficulty: multi-step security work in realistic codebases, closer to what a professional auditor actually does. ExploitGym is the sharpest instrument of the three — it measures progress on turning knowledge of a vulnerability into a working exploit, the offensive end of the craft. A model scoring 88, 71 and 55 per cent respectively across those tests is not merely answering questions about security; on the published evidence it can do substantial parts of the job. That is simultaneously the product pitch and the risk assessment, and it is why the same announcement that boasts the scores wraps the model in a vetting programme. Readers should still hold the numbers at arm's length: launch benchmarks are selected by the vendor, run under the vendor's conditions, and the history of the field is that independent replication trims most headline claims. What is not in doubt is the direction — each model generation is absorbing more of the security workflow, defensive and offensive alike.

The race Carbon enters

Carbon lands in the most crowded frontier contest yet. The Intelligence Index table Google cites has four systems within nine points: Carbon 59, Claude Opus 5.5 at 58, GPT-6 "Astra" at 57 and Argon at 50 — a spread small enough that the ranking lead is a marketing asset more than a capability chasm. The real contest is elsewhere: in agentic endurance (how long a model can work unsupervised without drifting), in price per completed task, and now in privileged capability tiers. Google's two-week cadence between Argon and Carbon also signals a release strategy of rapid, specialised drops rather than single annual flagships — a tempo that compresses rivals' response windows and keeps the company's models permanently in the launch-news cycle.

Pricing and availability

For those who do get access, Google has published aggressive introductory pricing: $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 per cent — a structure aimed squarely at long-running agentic workloads, where the same context is read repeatedly. The standard rates after the introductory period are listed at $4 and $20 per million tokens respectively. Even at standard rates, the pricing undercuts the assumption that frontier capability must keep getting more expensive per task, and it will pressure rivals whose top-tier models anchor the market's price ceiling.

Why it matters beyond the labs

Three things make Carbon a story beyond the AI industry. First, the normalisation of gated releases: the era of a frontier model simply appearing on an API for anyone with a credit card is ending for the most capable systems, replaced by vetting regimes that look increasingly like the export-control world of sensitive technology. Second, the defensive-offensive race is now explicit — Google is marketing a security model on its exploit-adjacent scores while restricting it for exactly that reason, a tension every lab now manages in public. Third, for developers and startups everywhere, including Pakistan's growing software-export sector, frontier coding capability at falling token prices changes what a small team can attempt: security audits, legacy-code modernisation and long-horizon automation that once needed large engineering organisations move within reach of firms that can get access — and access tiers, decided in California, become a quiet form of industrial policy.

What to watch next: the pace of Fairwind approvals, independent benchmark replication, and whether Carbon's general release follows Argon's two-week cadence or stretches as the security review deepens. TaazaPulse covers technology on the International desk; readers interested in how global platform shifts reach Pakistan can also read our explainer on the dollar rate and the Pakistani economy.