AiPhreaks ← Back to News Feed

Gemini 4 Carbon: Our next frontier model

By Jakub Antkiewicz

•

2026-10-29T14:57:57Z

Google DeepMind has announced Gemini 4 Carbon, its latest frontier model, just two weeks after the release of its predecessor, Gemini 4 Argon. The new model is positioned for its advanced capabilities in cybersecurity and long-horizon agentic tasks, immediately securing the top rank on the v5.0 Artificial Analysis Intelligence Index with a score of 59. This rapid release and high performance indicate a focus on specialized, high-stakes applications like software security and complex engineering problems.

Key Capabilities and Performance Benchmarks

Gemini 4 Carbon demonstrates significant performance gains across several industry benchmarks, particularly in the cybersecurity defense lifecycle. Google reports the model is built for sustained, multi-day workflows and will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Its benchmark achievements include:

  • SEC-bench Pro: Achieves 88.2% for autonomously finding vulnerabilities in real-world software.
  • ExploitGym: Scores a new state-of-the-art 54.8% for validating exploitable vulnerabilities.
  • CWE-bench: Reaches 71.0% for patching and fixing the root cause of vulnerabilities.
  • WebDev Arena: Jumps to #4 with a 1768 elo rating, showing improved quality in full-stack web development.

Phased Rollout and Enhanced Safety

Google is adopting a cautious deployment strategy, initially providing Gemini 4 Carbon access to a select group of cyber defenders and trusted partners via its Fairwind Program. This deliberate rollout is designed to gather feedback and refine safety guardrails before a wider public release. The model also incorporates notable safety enhancements over Argon, including achieving a 0.0% attack success rate on the Indirect Prompt Injection benchmark and reducing unsupported claims by approximately half in internal testing, leading to more reliable and honest interactions.

Google's rapid iteration from Argon to Carbon highlights a strategic focus on enterprise-grade security and reliability, positioning Gemini not just as a general-purpose model but as a specialized tool for critical industries where trust is paramount.
End of Transmission
Scan All Nodes Access Archive