Google Unveils Gemini 4 Carbon With Major Breakthroughs in Cybersecurity and Agentic Tasks

Google has officially released Gemini 4 Carbon, its newest frontier artificial intelligence model designed to deliver substantial advances in coding, autonomous agentic operations, and cyberdefense. The model arrives as an upgrade to Argon, which the company introduced earlier this month.

According to Google, Gemini 4 Carbon represents a major push forward in frontier AI capabilities, particularly in discovering security vulnerabilities and writing secure code.

Leading Benchmark Performance

Gemini 4 Carbon has captured the top position on the Artificial Analysis Intelligence Index (v5.0), ranking ahead of competing frontier models including Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.

The model also demonstrated dominant capabilities across specialized cybersecurity benchmarks:

• SEC-bench Pro: Holds the number one spot for identifying real-world software vulnerabilities, outperforming GPT-6 Astra.
• ExploitGym: Ranks first in testing an AI model's ability to turn identified vulnerabilities into functional exploits.

Multi-Day Persistence and Long-Horizon Tasks

Beyond security evaluation, Google built Gemini 4 Carbon to handle complex, long-horizon workflows. Internal teams at Google reported observing the model run continuously for days, working through dead ends to solve sophisticated, multi-step engineering assignments.

Google notes that the model shows enhanced resistance to prompt injection attacks, avoids taking shortcuts during problem-solving, and collaborates more effectively with users. It also improves reliability: Gemini 4 Carbon cuts unsupported claims by half compared to the Argon model.

Pricing and Rollout Plans

Access to Gemini 4 Carbon is initially restricted to a select group of cybersecurity defense firms. Availability will later widen to additional partners through Google's Fairwind Program.

Google is offering the model at the same introductory rate as its predecessor, priced at $2 per million input tokens and $10 per million output tokens.