Google Gemini 4 Carbon Leads Cybersecurity Benchmarks With Long Horizon AI
Google's Gemini 4 Carbon tops AI and cyber benchmarks, enters a phased defender rollout, and starts at $2 input and $10 output per million tokens at launch.
Summary
Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon. Google DeepMind SVP and Google chief AI architect Koray Kavukcuoglu calls it the company’s strongest model. Carbon ranks first on Artificial Analysis Intelligence Index v5.0 with 59, scores 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on Collinear’s CWE-bench for patching, all evaluated without guardrails. It rose 89 Elo points to fourth on WebDev Arena with 1,768 and can sustain agentic engineering work for days across coding, science, full-stack development and multimodal tasks.
Internal tests show more than 2x Argon’s secure-code success, while Chrome found Carbon produced high-quality patches over 10x better than another lab’s frontier model; Wiz also recorded leading web penetration testing results. Terminal-Bench 4.0 cheating fell to 0.0%, attempts to 0.4%, severe misalignment alerts nearly 80%, unsupported claims about 2x, overreach 26% and under-service 29%. Carbon resisted every Gray Swan indirect prompt-injection attempt, reaching 0.0% attack success versus Argon’s 0.7%. Google is submitting Carbon to voluntary U.S. government pre-release testing, initially limiting access to selected cyber defenders before adding trusted Fairwind Program partners, then developers, businesses and consumers. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper; later prices double to $4 and $20.
Positives
- Carbon ranks first on Artificial Analysis Intelligence Index v5.0 with a score of 59.
- SEC-bench Pro, ExploitGym and CWE-bench scores reached 88.2%, 54.8% and 71.0%, respectively.
- Chrome measured more than 10x better high-quality patch generation than another lab’s large frontier model.
- Terminal-Bench 4.0 successful cheating fell to 0.0%, while severe internal misalignment alerts declined nearly 80%.
- Gray Swan indirect prompt-injection attack success reached 0.0%, improving on Argon’s 0.7%.
- Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens.
Risks & concerns
- Carbon initially remains restricted to selected cyber defenders, with wider availability promised only as soon as possible.
- The external cybersecurity benchmark results were measured without the guardrails intended for deployed use.
- Several major performance claims rely on Google’s internal benchmarks and observations from Chrome and Wiz.
- Input and output prices double to $4 and $20 per million tokens after the introductory period.
- Carbon can run for days on difficult engineering workflows, indicating some long-horizon tasks may remain time-intensive.