Google DeepMind Previews Gemini 4 Carbon, Claiming Leadership in Cybersecurity and Long-Horizon Capabilities

Google has "pre"-released another model, the strongest Gemini 4?

Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 (score 59, with Claude Opus 5.5 at 58), and claims leadership on defensive benchmarks such as vulnerability discovery, exploit validation, and patch remediation; but it is not yet available for use—it is first being given to some cybersecurity defenders and trusted partners (the Fairwind program), and is participating in the U.S. government's voluntary pre-release testing, before gradually opening up to developers, enterprises, and consumers. https://t.co/N3yQOIrnZG

Capability narrative: harder, longer engineering tasks Carbon targets harder engineering problems that span longer time horizons, suited to sustained agentic workflows. Internal teams say it can run for days and keep making progress after getting stuck, for example locating root causes and applying patches in large codebases and low-level systems. Feedback-driven improvements cover coding, agents, scientific tasks, full-stack web, and multimodal understanding, among others. On web development, it is said that after output quality improvements, WebDev Arena rose by about +89 Elo, ranking 4th (about 1768 points).

Cybersecurity Officially, it claims a significant leap across the entire chain of "discovery → validating exploitability → fixing root causes"; all external benchmarks are measured without guardrails: · SEC-bench Pro about 88.2% (autonomously finding complex vulnerabilities in real software) · ExploitGym about 54.8% (judging whether a vulnerability is exploitable, to help prioritize) · Collinear CWE-bench about 71.0% (turning validated findings into working patches)

Internal metrics also include: compared with Argon, the success rate of generating secure code is claimed to improve by more than about 2x; the Chrome team says its high-quality patch performance is more than about 10x better than "another lab's large frontier model"; cloud security company Wiz also says it leads on web penetration testing-related benchmarks (related to Wiz's Scan for Good). These are mostly partner or internal observations, and the methodology needs separate verification.

Safety and alignment It continues Argon's multi-layer safety stack (training, real-time abuse protection, prompt injection resilience, chain-of-thought monitoring, etc.), and claims a further step forward on alignment, especially on high-pressure/complex tasks of "doing what should be done": · On Terminal-Bench 4.0, successful cheating dropped to 0.0%, cheating attempts were about 0.4%, while overall accuracy still improved; in internal live coding, high-severity misalignment alerts dropped by about 80%. · Compared with Argon, unsupported assertions are about halved; it more often writes methodological caveats into deliverables, or explicitly lists what evidence is missing and why it is not acting yet. When encountering ambiguity, it tends to first pull sources and surface contradictions to the user before acting. · Unauthorized actions dropped by about 26%, and under-service dropped by about 29%; multi-step plans are more likely to be completed without overstepping. · Gray Swan indirect prompt injection (IPI) benchmark: Argon's attack success rate is about 0.7%, while Carbon claims 0.0%.

Release and pricing Starting this week, it is being opened to some cybersecurity defenders, then gradually expanded to more trusted partners through Fairwind; after collecting feedback and iterating on guardrails, it will reach product surfaces as soon as possible. Introductory pricing is the same as Argon: about $2 per million input tokens, about $10 per million output tokens, and cached input at about a 95% discount off the input price. After the introductory period, it will become about $4 input and $20 output (per million tokens).