Google unveils Gemini 4 Carbon, its most powerful model, focused on cybersecurity and long-horizon tasks

Google has announced Gemini 4 Carbon, its new frontier model, which it describes as its most powerful yet, with leading capabilities in cybersecurity and “long-horizon” tasks. The announcement is signed by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect, and is dated October 28, 2026. Carbon arrives just two weeks after Gemini 4 Argon, which Google presents as evidence of the pace at which it is pushing the frontier.

According to the company, Carbon ranks first on the newly updated version 5.0 of the Artificial Analysis Intelligence Index, with 59 points. Google also claims notable improvements in coding and agentic tasks, as well as in the key capabilities for cybersecurity defense.

On external cybersecurity benchmarks, evaluated without safeguards, Google cites three results: 88.2% on SEC-bench Pro (vulnerability detection), 54.8% on ExploitGym (exploitation) and 71.0% on Collinear's CWE-bench (patching). It presents them as clear leadership at every stage of the defensive cycle: finding real flaws, validating their risk through exploitation and fixing the root cause.

In internal tests, Google says Carbon achieves more than double Argon's success rate at generating secure code. It adds that the Chrome team reported Carbon performed more than 10 times better than another lab's frontier model at generating high-quality patches. Wiz, Google's cloud security platform, also finds that Carbon leads in web penetration testing, a core capability of its Scan for Good initiative. These figures come from the company itself and from teams linked to it, not from independent verification.

In coding and reasoning, Google says Carbon is designed for long engineering problems and sustained agentic workflows. Its teams have seen it work for days, getting past dead ends, on tasks such as locating and fixing subtle bugs in large codebases and low-level systems. On WebDev Arena it climbs 89 Elo points to fourth place, at 1768.

On the model's own safety and alignment, Carbon relies on the same multilayered stack as Argon: model training, real-time defenses against misuse, resistance to prompt injection and chain-of-thought monitoring. Google says attempted cheating on Terminal-Bench 4.0 drops to 0.4% and successful cheating falls to 0.0%, with better overall accuracy. In real-world coding environments with internal teams, high-severity misalignment alerts fell by nearly 80%.

The company adds that Carbon roughly halves unsupported claims compared with Argon, and that when faced with ambiguity it retrieves sources and flags discrepancies to the user before acting. According to Google, it reduces overreach by 26% and underservice by 29%, which would make execution of multi-step plans more reliable. On Gray Swan's indirect prompt injection benchmark, the attack success rate drops from Argon's 0.7% to 0.0%.

The rollout will be gradual. Google says it is taking part in the US government's voluntary pre-release testing process. Starting this week, Carbon will be available to a subset of cybersecurity defenders, with access expanding to more trusted partners through the Fairwind Program. The company is still gathering feedback and adjusting safeguards before offering it to developers, businesses and consumers “as soon as possible,” without giving a date.

On pricing, the launch will keep Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. After the introductory period, prices will be $4 and $20 per million tokens, respectively.

The figures should be read with caution: nearly all come from Google itself and from internal testing. Even so, the message is clear: cybersecurity is becoming a central axis of competition among frontier models, with initial access restricted to defenders and trusted partners.