Google Unveils Gemini 4 Carbon, a Cyber-Focused Model It Is Releasing Slowly
Google's new Gemini 4 Carbon tops one popular AI ranking and posts strong security-testing scores, but only a small group of cyber defenders can use it for now.
Google announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon. It calls Carbon its strongest model so far, with particular strengths in cybersecurity and “long horizon” work, meaning tasks that take many steps over a long time without a human steering. The catch: for now, access is limited to a group of cyber defenders and then trusted partners through Google’s Fairwind Program. Google says developers, businesses and consumers will follow “as soon as possible”, but gives no date.
The numbers come from Google and partners, so read them as claims. Google says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. On security tests it reports 88.2% on SEC-bench Pro (finding software flaws), 54.8% on ExploitGym (turning flaws into working attacks) and 71.0% on CWE-bench (fixing them). It also says Carbon more than doubles Argon’s success at writing secure code, and that internal teams have watched it work for days on bugs in large codebases. Pricing is meant to match Argon: $2 per million input tokens and $10 per million output tokens at first, rising to $4 and $20 later. Tokens are the small chunks of text a model reads and writes.
Google also stresses safety work. It reports zero successful cheating on Terminal-Bench 4.0, about 80% fewer high-severity misalignment alerts in internal coding tests, and no successful attacks on the Gray Swan prompt-injection benchmark. Prompt injection means hiding instructions in content so an AI follows them instead of yours. It is also taking part in the US government’s voluntary pre-release testing.
What’s behind it: a model that is very good at finding flaws is also good at exploiting them. That is why Google is rolling it out to defenders first, a pattern we have seen with other labs’ security-capable models. Starting small gives the company time to watch for misuse before millions of people get access. The benchmark focus on security also signals where labs see the next big commercial market.
What this means for you: for most people, nothing yet. You cannot use Carbon today, and the context window size and usage limits were not published. If you are a developer or run a small business, watch for general availability and the price change after the introductory period. If you work in security, the Fairwind Program is the door to knock on. Worth keeping expectations grounded until independent testers have had their turn, especially since a tougher version of the Artificial Analysis index changed rankings elsewhere today.
Sources
Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Nvidia Releases Nemotron 3.5 Transcribe: One Open Model for Transcripts, Translation and Who Said What
Nvidia's new open speech model transcribes 40 languages, translates 31 and labels speakers in 13, and runs on a single GPU, though only for recorded audio.
