Google Gemini 4 Carbon Release Notes

Overview of Gemini 4 Carbon

Gemini 4 Carbon is Google DeepMind's latest frontier model, designed to excel in complex, long-horizon engineering tasks and cybersecurity defense. It ranks #1 on the Artificial Analysis Intelligence Index (v5.0) and demonstrates significant performance gains in coding, agentic workflows, and vulnerability management compared to its predecessor, Gemini 4 Argon.

Cybersecurity and Vulnerability Management

Carbon provides advanced capabilities across the entire cybersecurity defense lifecycle, including detection, validation, and remediation. The model achieves state-of-the-art results on key industry benchmarks:

  • Vulnerability Detection: 88.2% on SEC-bench Pro.
  • Exploitation Validation: 54.8% on ExploitGym.
  • Patching: 71.0% on CWE-bench.

Internal testing by the Chrome team indicated that Carbon performs over 10x better than other large frontier models when generating high-quality patches. Additionally, the model has demonstrated a 2x uplift in generating secure code compared to Argon.

Long-Horizon Agentic Capabilities

Carbon is engineered for sustained, multi-step engineering workflows. It is capable of operating over extended periods to resolve complex bugs and perform root-cause analysis across large codebases. In web development, the model achieved a #4 ranking on the WebDev Arena leaderboard, gaining 89 Elo points to reach a score of 1769. The model is designed to reduce overreach and under-service, improving the reliability of end-to-end plan execution.

Safety, Alignment, and Security

DeepMind has implemented a multi-layered safety stack for Carbon, focusing on alignment, transparency, and resilience against adversarial attacks:

  • Prompt Injection: Carbon achieved a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection (IPI) Benchmark.
  • Misalignment Reduction: High-severity misalignment alerts in internal coding environments dropped by nearly 80%.
  • Honesty: The model reduces unsupported claims by approximately 2x compared to Argon by surfacing methodological caveats or withholding actions when evidence is missing.
  • Shortcut Mitigation: Successful cheating on Terminal-Bench 4.0 was reduced to 0.0%.

Availability and Pricing

Google is adopting a phased rollout strategy for Gemini 4 Carbon. The model is currently being made available to a subset of cyber defenders and will be expanded to trusted partners via the Fairwind Program. A broader release for developers, businesses, and consumers is planned for the near future.

Carbon is priced at $2 per million input tokens and $10 per million output tokens, with cached input tokens available at a 95% discount.

Sources