Google DeepMind has introduced Gemini 4 Carbon, its latest frontier AI model, just two weeks after releasing Gemini 4 Argon. Announced on October 28, 2026, Carbon is positioned as the company’s strongest model to date, with particular strengths in cybersecurity defense and long-horizon agentic tasks.

A Deliberate Rollout for Cyber Defenders

Unlike a full public launch, Carbon is initially available to a subset of trusted cybersecurity defenders through Google’s Fairwind Program. Google is participating in the U.S. government’s voluntary pre-release testing process and plans to expand access gradually after gathering feedback on guardrails. The company aims to bring the model to developers, businesses, and consumers as soon as possible.

Pricing matches the introductory rates of Argon: $2 per million input tokens and $10 per million output tokens, with cached inputs at a 95% discount. After the introductory period, prices rise to $4 and $20 respectively.

Leading Performance on Intelligence and Cyber Benchmarks

Carbon ranks number one on the updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. It demonstrates significant gains in coding, agentic workflows, and scientific tasks. On the WebDev Arena leaderboard, it scores 1768, placing fourth.

In cybersecurity, the model sets new state-of-the-art results across the defense lifecycle:

  • Vulnerability detection: 88.2% on SEC-bench Pro, autonomously finding complex flaws in real-world software.
  • Exploitation validation: 54.8% on ExploitGym, helping prioritize truly exploitable issues.
  • Patching: 71.0% on Collinear’s CWE-bench, generating working root-cause fixes.

Internal benchmarks show more than a 2x improvement in success rate for generating secure code compared to Argon. Google’s Chrome team reported that Carbon performed more than 10 times better than another lab’s frontier model when producing high-quality patches. Wiz observed leading performance on web penetration testing benchmarks.

Long-Horizon Agentic Capabilities

Carbon is designed for sustained, multi-step engineering work. Internal teams have used it to run for days on complex tasks such as root-causing and patching subtle bugs across large codebases and low-level systems. It shows improved quality in full-stack web development and multimodal understanding, with better artifact generation.

Advances in Safety and Alignment

The model builds on Argon’s multi-layered safety stack, including training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. Key improvements include:

  • Reduced shortcut exploitation: successful cheating on Terminal-Bench 4.0 dropped to 0.0%, with attempts at 0.4%.
  • Greater honesty: unsupported claims cut by approximately 2x versus Argon.
  • Better collaboration under ambiguity: the model surfaces discrepancies and retrieves sources before acting.
  • Balanced execution: overreach reduced by 26% and under-service by 29%.
  • Prompt-injection resistance: 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark, down from Argon’s 0.7%.

High-severity misalignment alerts in live coding environments fell by almost 80%.

Why This Matters

Gemini 4 Carbon highlights the accelerating pace of frontier model development, with Google shipping a stronger system only weeks after Argon. Its focus on cybersecurity tools for defenders comes as AI is increasingly used on both sides of security operations. For developers and security teams, the model’s long-horizon reliability and reduced fabrication rates could improve automated code review, vulnerability management, and patch generation. The staged rollout reflects ongoing caution around powerful coding and cyber capabilities. As access expands, organizations will be able to evaluate how these gains translate to production workflows.

Sources

[1] Google DeepMind: Introducing Gemini 4 Carbon [2] Reuters: Google launches Gemini 4 Carbon AI model with cybersecurity focus