Google Rolls Out Gemini 4 Carbon to Cyber Defenders, Tops AI Rankings
Google’s newest frontier model, Gemini 4 Carbon, leads the Artificial Analysis Intelligence Index and is now available to a limited group of cyber defenders.
Launch and Availability
Google announced Gemini 4 Carbon, its latest frontier AI model, in a blog post on Wednesday. The announcement came from Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect. Carbon is being delivered to a small group of cyber defenders, following a similar rollout pattern used for Gemini 4 Argon.
Starting this week, the model will be provided to a subset of trusted partners through Google’s Fairwind Program, with plans to expand to additional collaborators. Google also intends to participate in the U.S. government’s voluntary testing process for AI models before broader release.
Performance Rankings
On the Artificial Analysis Intelligence Index v5.0, Carbon earned a top score of 59, placing it ahead of Anthropic’s Claude Opus 5.5 (58) and OpenAI’s GPT‑6 Astra (57). The index also ranks Carbon second on its Coding Agent Index, behind Claude Code running Opus 5.5.
In a separate knowledge test, Carbon answered 58 % of questions correctly, an improvement over Argon’s 48 %. However, its hallucination rate on that test rose from 4 % to 11 %.
Cybersecurity Benchmark Results
Google reports that Carbon leads three external cybersecurity tests run without guardrails. It achieved an 88.2 % score on SEC‑bench Pro, which measures the ability to find real vulnerabilities. On ExploitGym, which evaluates exploitability, Carbon scored 54.8 %, and on CWE‑bench, which assesses patching, it earned 71.0 %.
Internal testing showed Carbon was more than twice as successful as Argon at writing secure code. Google’s Chrome team found it over ten times better than another lab’s large frontier model at producing high‑quality patches, though the lab was not named.
Reliability and Misalignment
Google states that Carbon generates roughly half as many unsupported claims as Argon. In internal coding environments, high‑severity misalignment alerts dropped by almost 80 %. On Gray Swan’s prompt‑injection benchmark, no evaluated attack succeeded against Carbon, compared with a 0.7 % success rate for Argon. Terminal‑Bench 4.0 also recorded a 0.0 % cheating success rate for Carbon.
Other Benchmark Highlights
Carbon did not lead in every test. GPT‑6 Astra outperformed it on FrontierSWE 2, a software‑engineering benchmark, with 65.5 % versus Carbon’s 61.2 %. Claude Opus 5.5 led on OSWorld, a computer‑use measure, scoring 87.2 % against Carbon’s 81.5 %.
On the WebDev Arena leaderboard, which ranks AI‑built websites, Carbon climbed 89 points to reach fourth place.
Pricing and Future Availability
Carbon will launch at Argon’s introductory price of $2 per million input tokens and $10 per million output tokens. A footnote in Google’s post indicates that prices are slated to increase to $4 and $20, though no timeline was provided.
Google did not release a technical report for Carbon at the time of reporting.
Corporate Context
The announcement coincided with Alphabet’s third‑quarter earnings release and followed DeepMind’s release of the Genie 4 world model. Sundar Pichai, CEO of Google and Alphabet, referenced Carbon during the earnings call.
Google’s selective rollout of Gemini 4 Carbon underscores a strategic focus on securing AI capabilities before broader market exposure.
