Google launches Gemini 3.8 Flash and Flash Cyber

On Wednesday, Google unveiled two iterations of its 3.8 Flash model. One variant, Flash, serves as a versatile "workhorse" for various tasks including software development and multi-step reasoning. The other, Flash Cyber, is a more specialized model crafted to detect and resolve vulnerabilities. Google CEO Sundar Pichai announced that 3.8 Flash provides substantial improvements over its predecessor, 3.7 Flash, across a range of areas.

For instance, it surpassed many frontier models on the DeepSWE coding benchmark, all while maintaining lower costs. Meanwhile, Flash Cyber is touted as Google's most potent cybersecurity model, matching frontier-level performance in vulnerability detection and patching. It scored 86.2% on the CyberGym cybersecurity benchmark and 47.2% on CWE-Bench, a measure of AI patching abilities.

Google's internal benchmark revealed that 3.8 Flash had a success rate of over 70% in identifying vulnerabilities across 20 programming languages. This third Flash release in six weeks follows the release of version 3.7. Google made 3.8 Flash available immediately in Gemini Enterprise, accessible via the Gemini API within various platforms such as Google AI Studio, Google Antigravity, Android Studio, and Stitch.

It is priced at $0.75 per million input tokens and $3.75 per million output tokens, mirroring the introductory pricing of Gemini 3.7 Flash. Users can adjust the model's effort levels based on their specific needs for quality, cost, and latency. 3.8 Flash can process a variety of data types, including text, images, audio, video, and PDF files.

It operates within a 1 million-token input window and a 64,000-token output limit, and can ingest and generate content in multiple formats. Benchmarks showcased 3.8 Flash's prowess in coding, multimodal capabilities, computer interaction, long-context tasks, scientific reasoning, and specialized domains like finance and law. For example, it outperformed its predecessor and other frontier models on finance-specific tasks and legal reasoning benchmarks, crafting a game and a DOS version of Google Maps in demonstration of its versatile skills.

In terms of cybersecurity, Flash Cyber is currently being trialed with "trusted defenders" through Google's Fairwind Program, targeting government authorities and critical infrastructure operators. This model is being trained rigorously in cybersecurity and has demonstrated a significant leap in prompt injection robustness, excelling in autonomous vulnerability discovery and patching.

It is particularly adept at securing Google's own code, generating significantly more correct patches compared to larger commercial models.