Google launches Gemini 4 Carbon for automated software security

A specialized tool for defenders

Google has introduced Gemini 4 Carbon, a new artificial intelligence model specifically engineered to handle complex cybersecurity and coding tasks. Unlike general-purpose models, Carbon is built to navigate the difficult process of finding software flaws, determining if those flaws can be exploited by attackers, and generating functional patches to fix them.

The model builds upon the previous Gemini 4 Argon architecture but is optimized for long-duration tasks that often take human engineers days to complete. It is designed to navigate complex paths to identify tiny errors hidden within massive amounts of code and deep system layers. This capability allows it to address the root causes of security issues rather than just treating the symptoms.

Access to the model will not be immediate for the general public. Instead, Google plans to begin the rollout with selected cyber defenders. These early users will help refine the tool before it moves to additional trusted partners through Google’s Fairwind Program.

Benchmark results and performance

To measure its effectiveness, Google tested Carbon across three linked stages of security work: discovery, exploitation, and patching. As reported by CyberSecurityNews, Carbon achieved a score of 88.2% on SEC-bench Pro for finding vulnerabilities. It also scored 54.8% on ExploitGym for testing exploitation and 71.0% on CWE-bench for creating patches.

In internal testing, Google reported that Carbon more than doubled the success rate of the Argon model when generating secure code. The Chrome team also noted that Carbon performed over tenfold better than an unnamed rival model in producing high-quality patches. While Google did not name the competitor, these figures highlight the model's specialized focus on code integrity.

The model currently sits at the top of the Artificial Analysis Intelligence Index v5.0 with a rating of 59. This index evaluates nine different categories, including science and general knowledge. While this provides a broad overview of the model's intelligence, Google notes that it does not replace the specific security testing required for production environments.

Safety layers and guardrails

Security is a primary focus for Carbon, and Google has integrated several layers to protect against misuse. These include specific defenses against prompt injections, which occur when hidden instructions are used to manipulate an AI. The model also includes monitoring systems to oversee its reasoning and actions during complex tasks.

In testing against the Gray Swan Indirect Prompt Injection benchmark, Carbon showed a 0.0% success rate for attacks. This was a notable improvement over the Argon model, which saw a 0.7% success rate on the same test. However, Google clarifies that these results apply only to the specific attacks evaluated and do not guarantee safety against every possible attack in a live environment.

Google also reported significant improvements in internal coding environments using the model. High-severity misalignment alerts fell by almost 80% during these tests. Additionally, unsupported claims dropped by roughly half compared to the Argon model, and successful cheating on Terminal-Bench 4.0 fell to zero.

Deployment and pricing structure

Before a wider release, Google is participating in voluntary testing with the U.S. government. This phase is intended to ensure the model meets necessary standards before it reaches a broader audience. Once the initial defenders are satisfied, the model will move through the Fairwind Program to other trusted partners.

The pricing for Carbon follows a tiered structure based on token usage. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens. To encourage efficient usage, Google is offering a 95% discount on cached inputs during this initial period.

Users should be aware that these introductory rates are temporary. Standard rates are scheduled to rise to $4 per million input tokens and $20 per million output tokens at a later date. While a firm date for general availability has not been announced, the pricing structure provides a clear roadmap for how the service will be monetized as it scales.