Google on Wednesday, October 28, announced Gemini 4 Carbon, a new frontier AI model that the company describes as its strongest yet, with a focus on cybersecurity and long-running tasks. The catch: almost nobody can try it. Access is limited to a small group of cyber defenders at first, so most of what the public knows comes from Google’s own announcement.
In a post on Google’s blog, Koray Kavukcuoglu of Google DeepMind says Carbon builds on Gemini 4 Argon, which launched about two weeks earlier. Ars Technica frames Carbon as a modest step up from Argon rather than a new generation, a useful reminder that a fast release cadence does not always mean a big leap.
What Google says it can do. Google reports that Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. On security benchmarks, it lists 88.2% on SEC-bench Pro, which tests finding vulnerabilities, 54.8% on ExploitGym, which tests exploitation, and 71.0% on CWE-bench, which tests patching. Google calls the ExploitGym result a new state of the art. On WebDev Arena, a coding leaderboard, it says Carbon gained 89 Elo points to reach 1768, ranking fourth.
Google also says Carbon more than doubles Argon’s success rate on its internal secure-code tests, and that its Chrome team saw it do far better than another lab’s large model on high-quality patches. Engadget and SiliconANGLE both report the headline scores, but they are company-reported, and Google notes the cybersecurity tests ran without the guardrails that would apply in real products.
Who can use it, and when. This week, Carbon is available to a subset of cyber defenders. Google says access will widen gradually to trusted partners through what it calls the Fairwind Program, and that developers, businesses and consumers will get it later, after more feedback and work on safeguards. It has not given a date.
What it costs. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Google says the price will rise to $4 and $20 per million tokens after the introductory period, which matters for anyone budgeting a product around it.
Safety claims. Google says Carbon uses the same layered safety approach as Argon and takes part in the US government’s voluntary pre-release model testing process. It also reports that a prompt-injection benchmark, Gray Swan IPI, showed a 0.0% attack success rate, compared with 0.7% for Argon, and that internal high-severity misalignment alerts fell by almost 80%. These are internal results and have not been independently confirmed in the post.
Why it matters. The same skills that help a defender find and patch a flaw can help an attacker exploit one. That dual-use risk is likely why Google is starting with security teams instead of a public launch. For readers in the US and UK, where businesses and governments are under constant pressure from ransomware and software supply-chain attacks, the question is whether controlled access gives defenders a head start, and how long it lasts.
What to watch. Look for independent testing once more researchers get access, a technical report with details Google has not yet published, and a date for broader availability. Until then, treat the benchmark numbers as a promising claim, not a verdict.
Sources: Google, Ars Technica, Engadget and SiliconANGLE. Figures are as reported by Google unless stated otherwise.
Googleplex headquarters in Mountain View, California (file image), Wikimedia Commons, CC BY-SA 4.0
