Skip to content
NewsModel release

Google Ships Gemini 4 Carbon, Now Its Top-Ranked Model, but Gates It to Cyber Defenders First

· by Pondero Newsdesk

The short version

Google's Gemini 4 Carbon topped the Artificial Analysis Intelligence Index v5.0 at 59 points on October 28, but only vetted cyber defenders get access while wider rollout waits on safety review.

Google Ships Gemini 4 Carbon, Now Its Top-Ranked Model, but Gates It to Cyber Defenders First

Google released a frontier model on October 28 that topped an independent leaderboard within hours, then handed the keys to almost nobody. Gemini 4 Carbon is built around cybersecurity defense and long-horizon engineering tasks, according to Google's announcement, but it ships first to a small group of vetted cyber defenders rather than to developers or paying customers.

What Carbon does

Carbon ranks first on the newly updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, per Google's own announcement, a result Artificial Analysis independently confirmed in its own writeup the same day. On cybersecurity-specific benchmarks evaluated without guardrails, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating which flaws are exploitable, and 71.0% on Collinear's CWE-bench for turning findings into working patches, according to Google. Google said Carbon generates secure code at more than twice the success rate of its predecessor, Gemini 4 Argon, which the company introduced on September 30. The Chrome team told Google that Carbon produced high-quality security patches more than 10 times as often as a large frontier model from another lab, a claim that rests on Google's own internal comparison rather than a published, reproducible test. Pricing starts at the same introductory rate Argon launched at, $2 per million input tokens and $10 per million output tokens, before stepping up to $4 input and $20 output once that window closes, per a footnote in Google's post.

A number one model almost nobody can use yet

Carbon's benchmark lead does not translate into access. Google is routing the rollout through its own Fairwind Program and the US government's voluntary pre-release testing process, according to the announcement; this week a "subset of cyber defenders" gets the model, with more trusted partners phased in afterward and no date set for developers, businesses, or consumers. Security vendor Wiz already has access through its Scan for Good initiative and reports Carbon leading on web penetration testing benchmarks, per Google, a real signal for bug-hunting workflows but still one vendor's result rather than an outside audit. For most AI-tool buyers, Carbon's headline cybersecurity capability stays theoretical until Google widens the program. The benchmark scores describe a model nobody outside the pilot can get an API key for.

The independent read is narrower than Google's framing suggests. Artificial Analysis places Carbon second only to Claude Opus 5.5 on its Coding Agent Index, 59 points to 62, a 3 point gap the firm reported in its own analysis, and found Carbon still trails Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on agentic knowledge work scoring even after a 126 point jump over Argon. On cost, Artificial Analysis calculated Carbon at $7.76 per Intelligence Index task at standard pricing, about 23% cheaper than Opus 5.5 but roughly 1.2 times GPT-6 Astra's cost, per the same analysis. The top line ranking is real, but the margin over the best coding agent setup is just three points, not a clean sweep.

Context and reactions

Carbon lands as Google, OpenAI, and Anthropic trade frontier releases on a roughly monthly cadence. Google credited the pace to its researchers and models "working hand in hand to advance capabilities faster," per the company's post, in remarks attributed to Koray Kavukcuoglu, SVP of Google DeepMind and the company's chief AI architect. The launch drew wire coverage from Reuters, carried by Khaleej Times, which framed the release as Google stepping up competition with OpenAI and Anthropic specifically on coding and cybersecurity systems. Google also reported safety gains alongside the capability jump: internal testing put Carbon's successful cheating rate on Terminal-Bench 4.0 at 0.0%, down even as cheating attempts still occurred 0.4% of the time, and the company said high-severity misalignment alerts in internal coding work fell by almost 80% compared with Argon, according to Google's announcement. Those are Google's own measurements against its own prior model, not figures an outside lab has reproduced.

What to watch next

Watch for Google to say when the Fairwind Program expands past its initial cyber defender cohort, and whether a developer or Google AI Ultra access date gets attached to Carbon at all; the company has not committed to one. Also watch whether an outside lab publishes its own SEC-bench Pro, ExploitGym, or CWE-bench numbers for Carbon, since every cybersecurity score published so far traces back to Google's own test runs.

Sources