Google gives Gemini 4 Carbon to cyber defenders first
Google says Gemini 4 Carbon leads tests of finding, exploiting, and patching software flaws, and wider access will follow as it refines guardrails.
In brief
- Gemini 4 Carbon goes first to cyber defenders, then Fairwind Program partners
- Google says Carbon tops Artificial Analysis Intelligence Index v5.0 with 59 points
- Gemini 4 Carbon price matches Argon at $2 and $10 per million input and output tokens
Google DeepMind released Gemini 4 Carbon on Wednesday, two weeks after Gemini 4 Argon, and said in the company's post that a subset of cyber defenders would get the model first.
The two launches in a month follow nearly a year in which Google released no new flagship model, The Information reported. Google called Carbon its strongest model, a frontier model, the most capable class of system a lab builds, and is widening access in stages because of its cybersecurity skills.
After the first cyber defenders, access will extend gradually to more trusted partners through Google's Fairwind Program. The company said it is taking part in the US government's voluntary process for testing models before release, and will offer Carbon to developers, businesses, and consumers after gathering feedback and refining guardrails, “as soon as possible.”
Google said Carbon ranks first, at 59 points, on version 5.0 of the Artificial Analysis Intelligence Index, an independent composite ranking of models. The Verge reported that Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra trail it on that index.
In security tests run without guardrails, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for turning them into working exploits, and 71.0% on Collinear's CWE-bench for writing fixes, Google said. The company also said Carbon more than doubled Argon's success rate at generating secure code in internal tests, and that its Chrome team found Carbon more than 10 times better at writing patches than another lab's large frontier model. These company-reported results have not been independently verified.
Evaluation platform Vals AI placed Carbon second on its index at 70.36% accuracy, behind Anthropic's Claude Sonnet 5.5 at 70.85%, in a post on X. Arena.ai put Carbon first in its Text Arena and second in its Agent Arena, behind Claude Opus 5.5, according to its post on X.
Google said Carbon cut high-severity misalignment alerts, flags that a model strayed from intended behavior, by almost 80% in internal coding work. On Gray Swan's indirect prompt injection benchmark, which tests resistance to malicious instructions hidden in content, no attack succeeded, compared with a 0.7% success rate for Argon, the company said.
Carbon will launch at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, the units of text models process, with cached input tokens priced 95% lower.
Google released Gemini 4 Argon earlier in October. Google DeepMind, the Alphabet unit that develops the Gemini models, has also been expanding its work with governments, and the UK AI Security Institute said on October 25 it was deepening its partnership with the lab.