Introducing Gemini 3.8 Flash and 3.8 Flash Cyber

Google DeepMind's rapid iteration in its Gemini Flash series culminates in the release of Gemini 3.8 Flash and its specialized variant, 3.8 Flash Cyber. This launch underscores a strategic push towards more intelligent, efficient, and domain-specific AI models. The core advancements stem from rigorous training, including extensive exposure to the demanding field of cybersecurity, which has yielded significant gains in both reasoning and coding capabilities across the shared foundational intelligence of both models. The company emphasizes that these models are designed to work harder on complex tasks, executing more reasoning steps and iteratively calling tools to maximize performance, even if it occasionally means using more tokens.

Gemini 3.8 Flash

Gemini 3.8 Flash is positioned as an intelligent workhorse model, demonstrating substantial improvements over its predecessor, 3.7 Flash. It excels in long-horizon software engineering, often matching or exceeding the performance of more expensive frontier models on benchmarks like DeepSWE v1.1. This model is also engineered for dependability in critical enterprise autonomy, showing superior performance in specialized knowledge domains such as finance and legal analysis, as evidenced by benchmarks like Vals Finance Agent V2 and Harvey's Legal Agent Benchmark. Its ability to handle multi-step reasoning across diverse fields, including STEM and humanities, is highlighted by a 54.9% score on HLE-Verified.

The enhanced performance of 3.8 Flash is attributed to its design philosophy of 'working harder,' which involves executing extra reasoning steps and iterative tool calls. While this approach maximizes performance, developers have the flexibility to adjust 'effort levels' to manage token overhead for applications where compute efficiency is paramount. Google DeepMind showcases the model's versatility through examples like building a 3D wizard game, a functional DOS version of Google Maps, and an interactive 3D visualizer for hardware anatomy, all from simple prompts within Google Antigravity or AI Studio.

CyberGym

Gemini 3.8 Flash Cyber distinguishes itself with frontier-level performance in cybersecurity, specifically in autonomous vulnerability discovery. On CyberGym, a standard industry benchmark for finding vulnerabilities, the model surpasses both its predecessor, 3.5 Flash Cyber, and significantly larger frontier models. This achievement is particularly noteworthy as CyberGym primarily focuses on C/C++ codebases, a critical area for software security.

Beyond C/C++, Google DeepMind also evaluated 3.8 Flash Cyber against an internal benchmark encompassing a wide range of vulnerabilities across complex codebases in 20 programming languages. In this more comprehensive real-world scenario, the model demonstrated an impressive leap, achieving a success rate exceeding 70%. This broad language support and high success rate underscore its potential to provide a decisive advantage to defenders in today's complex and multi-faceted cybersecurity landscape.

Fairwind Program

Gemini 3.8 Flash Cyber is made available to trusted defenders through a new initiative called the Fairwind Program. This program is designed to equip cybersecurity professionals with expert capabilities, focusing on defensive actions like vulnerability fixing rather than offensive exploitation. The emphasis on automated patching is a key differentiator, with the model demonstrating strong performance on challenging external benchmarks like CWE-Bench, run by Collinear.

On CWE-Bench, Gemini 3.8 Flash Cyber achieves a pass@1 score of 47.2%, closely matching a leading frontier model at 47.8%, but at a significantly lower cost. This cost-effectiveness, combined with its speed, enables quick iteration in defensive strategies. Google itself is already leveraging 3.8 Flash Cyber to secure its own code, with the Chrome Security team reporting that the model produced 2.6 times more correct patches for Chrome vulnerabilities compared to the best commercial models, which are often much larger.