Google DeepMind Unveils Gemini 4 Carbon: A Frontier Model for Cyber Defense and Agentic Coding
Google DeepMind has introduced Gemini 4 Carbon, its most capable model to date. Carbon is a frontier model designed for long-horizon agentic coding and cybersecurity defense. It ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. For developers, however, access is the catch: Carbon is initially available only to a vetted group of cyber defenders, not the public API.
TL;DR
- Size: Parameter count and context window are not disclosed. Its predecessor, Gemini 4 Argon, supports a 1,048,576-token context.
- Deployment: Google-hosted only. No open weights, no self-hosting, and no quantized variants.
- Performance: Leads 15 of 17 published benchmarks against Argon, GPT-6 Astra, Opus 5.5, and Fable 5.1.
- Best result: 96.2% BFS F1 on GraphWalks (256K to 1M tokens), versus 84.2% for Argon.
- Weakest result: 81.5% on OSWorld-2.1, trailing Claude Opus 5.5 at 87.2%.
- Bottom line: State-of-the-art vulnerability finding, exploitation, and patching—but you probably cannot call it yet.
What Is Gemini 4 Carbon?
Carbon is the second Gemini 4 model, following Gemini 4 Argon. It is built specifically for high-stakes, long-horizon workflows in cybersecurity and agentic software development. The model is designed to identify vulnerabilities, reason about exploitation paths, and generate patches within complex codebases.
Benchmark Performance
Carbon leads 15 of 17 published benchmarks against competitive frontier models, including GPT-6 Astra, Claude Opus 5.5, and Fable 5.1. Its strongest showing is on GraphWalks, where it achieves 96.2% BFS F1 across context lengths from 256K to 1M tokens—a notable improvement over Argon's 84.2%. Its weakest result is on OSWorld-2.1, where it scores 81.5%, behind Claude Opus 5.5 at 87.2%.
Access and Availability
Carbon is available only through Google-hosted infrastructure. There are no open weights, no self-hosting options, and no quantized variants. Access begins with a vetted group of cyber defenders rather than the general public API.
The 2026 Context
Carbon's release reflects a broader 2026 trend: frontier AI labs are shipping domain-specialized models with restricted access for sensitive use cases such as cyber defense. As agentic coding and security workloads mature, vetted-access programs are becoming a standard distribution model for the most capable systems.
via MarkTechPost