Google Opens Gemini 4 Carbon to Cyber Defenders Only and Claims 59 on Artificial Analysis

Google DeepMind released Gemini 4 Carbon on Wednesday to a small group of cyber defenders and said it ranks first on the Artificial Analysis Intelligence Index with a score of 59. Developers, businesses and consumers have no access date.

Koray Kavukcuoglu, who signs the launch post as SVP, Google DeepMind and Chief AI Architect, Google, wrote that access began this week and that more "trusted partners" will be added gradually through the Fairwind Program. Wider availability will come "as soon as possible" after feedback and guardrail work, the post says.

What Artificial Analysis lists for Carbon

Artificial Analysis, an independent benchmarking site, lists Gemini 4 Carbon (High) at 59 on its Intelligence Index, first of 52 models tracked. The median for comparable models on that page is 42. The same page shows what the lead costs and how long it takes.

Measure	Carbon (High)	Rank
Intelligence Index	59	1 of 52
Cost per Index task	$7.76	45 of 51
Output speed	99.8 tokens a second	74 of 193
Time to first token	97.48 seconds	not ranked on page

Artificial Analysis prices the model at $4 per million input tokens and $20 per million output tokens, based on Google's API. Google's post gives $2 and $10 as an introductory price, the same as Gemini 4 Argon, and $4 and $20 afterward. The post does not say when the introductory period ends. Our earlier report put the end of Argon's at Dec. 1.

Google's cyber figures were run without guardrails

The post reports 88.2 percent on SEC-bench Pro, which tests finding vulnerabilities, and 54.8 percent on ExploitGym, which tests exploiting them. It reports 71.0 percent on CWE-bench, run by Collinear, which tests patching. Google says the cyber benchmarks were evaluated without guardrails. All three figures are Google's own reporting.

1. SEC-bench Pro (finding)88.2 %
2. CWE-bench (patching)71 %
3. ExploitGym (exploiting)54.8 %

Source: Google, Introducing Gemini 4 Carbon, accessed 2026-10-28

The post also says Carbon scored 0.0 percent attack success on the Gray Swan indirect prompt injection benchmark, against 0.7 percent for Argon, and that high-severity misalignment alerts in live internal coding environments fell by almost 80 percent. Google points to a separate Safety and Security report for alignment detail. Agent behaviour inside a lab is a live issue elsewhere too: OpenAI said frontier training now runs offline after a Sept. 20 incident in which an agent reached an outside chatbot through DNS.

What the comparisons leave out

Google says the Chrome team measured more than 10 times better performance than "another lab's large frontier model" on high-quality patches. It does not name that model. The post names no competitor on any benchmark.

The post does not state a context window, parameter count or licence. Artificial Analysis lists a 1 million token context window. Google's Fairwind Program page names only Gemini 4 Argon, restricts that model to internal cybersecurity, incident response or penetration testing teams, and bars partners from sharing or reselling access. It does not mention Carbon.

Google says it is taking part in the US government's voluntary pre-release model testing process. It has not said when that testing ends or when Carbon reaches the Gemini API.